priming sentence planning - mpg.pure

Cognitive Psychology 73 (2014) 1–40

Contents lists available at ScienceDirect

Cognitive Psychology

journal homepage: www.elsevier .com/locate/cogpsych

Priming sentence planning

http://dx.doi.org/10.1016/j.cogpsych.2014.04.0010010-0285/� 2014 Elsevier Inc. All rights reserved.

⇑ Corresponding author. Address: Max Planck Institute for Psycholinguistics, PO Box 310, 6500 AH NThe Netherlands.

E-mail address: [email protected] (A.E. Konopka).

Agnieszka E. Konopka ⇑, Antje S. MeyerMax Planck Institute for Psycholinguistics, Radboud Universiteit Nijmegen, The Netherlands

a r t i c l e i n f o

Article history:Accepted 8 April 2014

Keywords:IncrementalityMessage and sentence formulationLexical primingStructural priming

a b s t r a c t

Sentence production requires mapping preverbal messages ontolinguistic structures. Because sentences are normally built incre-mentally, the information encoded in a sentence-initial incrementis critical for explaining how the mapping process starts and forpredicting its timecourse. Two experiments tested whether andwhen speakers prioritize encoding of different types of informationat the outset of formulation by comparing production of descrip-tions of transitive events (e.g., A dog is chasing the mailman) thatdiffered on two dimensions: the ease of naming individual charac-ters and the ease of apprehending the event gist (i.e., encoding therelational structure of the event). To additionally manipulate easeof encoding, speakers described the target events after receivinglexical primes (facilitating naming; Experiment 1) or structuralprimes (facilitating generation of a linguistic structure; Experiment2). Both properties of the pictured events and both types of primesinfluenced the form of target descriptions and the timecourse offormulation: character-specific variables increased the probabilityof speakers encoding one character with priority at the outset offormulation, while the ease of encoding event gist and of generat-ing a syntactic structure increased the likelihood of early encodingof information about both characters. The results show that formu-lation is flexible and highlight some of the conditions under whichspeakers might employ different planning strategies.

� 2014 Elsevier Inc. All rights reserved.

ijmegen,

http://crossmark.crossref.org/dialog/?doi=10.1016/j.cogpsych.2014.04.001&domain=pdf

http://dx.doi.org/10.1016/j.cogpsych.2014.04.001

mailto:[email protected]

http://dx.doi.org/10.1016/j.cogpsych.2014.04.001

http://www.sciencedirect.com/science/journal/00100285

http://www.elsevier.com/locate/cogpsych

2 A.E. Konopka, A.S. Meyer / Cognitive Psychology 73 (2014) 1–40

1. Introduction

To produce language, speakers must decide what they want to say and how they want to say it –that is, they must formulate a preverbal message and a corresponding utterance. At the sentence level,the formulation process involves several steps. For example, when asked to describe a picture of a dogchasing a mailman, speakers must select referential terms from a range of potentially suitable nouns(e.g., man or mailman to refer to the patient in this event) and must select one out of a range of suitablesyntactic structures (e.g., active, passive, or intransitive constructions). Numerous production studieshave shown that the availability of lexical and structural information can influence selection processesas well as production speed (e.g., Bock, 1986a, 1986b; Smith & Wheeldon, 2001). Questions about therelative contributions of words and structures to grammatical encoding have inspired a number ofhypotheses about interactions between these processes (Bock, 1982; Bock & Griffin, 2000;Hartsuiker, Bernolet, Schoonbaert, Speybroeck, & Vanderelst, 2008; Pickering & Branigan, 1998) andhave led to the development of detailed production models (e.g., Chang, Dell, & Bock, 2006;Kempen & Hoenkamp, 1987).

Differences between models reflect different assumptions about the division of labor between lex-ical and structural processes in the shaping of sentence form (Bock, 1987a). On the one hand, lexicalistaccounts propose that structure building has a lexical source (e.g., Bates & MacWhinney, 1982):retrieving a word provides access to structural information stored with this word at the lemma leveland thus triggers the assembly of a syntactic structure. On the other hand, abstract structural accountsposit that structures can also be built by lexically-independent structural procedures (Bock, 1986a):when preparing their utterances, speakers may first generate an abstract structural framework andthen retrieve the necessary words in the order required by these structures. Experimental work testingthese accounts is found in the production (Bock, 1986a, 1986b, and others) as well as acquisition(Fisher, 2002; Tomasello, 2000) literature.

Here we take the position that debates about the relative timing of lexical and structural encodingare also important for explaining how speakers formulate and map preverbal messages onto language.Namely, production processes can be divided into two large classes, one concerned with encoding ofindividual elements of a message (non-relational processes) and the other concerned with encodingthe relationships between them (relational processes). The distinction applies both to sentence-leveland message-level encoding. At the sentence level, non-relational and relational information is carriedby words and structures respectively; at the message level, these processes refer to identification ofcharacters participating in an event (a dog, a mailman) and to encoding of the who-did-what-to-whom,relational structure of the event (one character chasing the other character). Since some combination ofnon-relational and relational processing at the message level and at the sentence level is necessary toproduce any utterance longer than one word, the coordination of these processes is important forexplaining information flow in the production system from conceptualization to linearization.

A crucial part of this puzzle is the fact that message-level and sentence-level processes are nor-mally interleaved during production. All psycholinguistic models agree that messages and sentencesare built incrementally, i.e., that speakers plan what they want to say in small chunks rather than insentence-sized units (Levelt, 1989; see Wheeldon, 2013, for a review). The high degree of temporaloverlap in message-level and sentence-level encoding requires a theory about dependencies betweenconceptual and linguistic processes. Notably, the two leading accounts of incrementality in sentenceproduction take different views on the way that speakers generate message-level and sentence-levelincrements. One proposal (linear incrementality; Gleitman, January, Nappa, & Trueswell, 2007)assumes that speakers can prepare a sequence of small conceptual and linguistic increments withoutguidance from a higher-level framework. The other proposal (hierarchical incrementality; Bock, Irwin,& Davidson, 2004; Bock, Irwin, Davidson, & Levelt, 2003) assumes that formulation can instead beginwith encoding of the gist of an event and with generation of a conceptual framework to guide subse-quent linguistic encoding. The difference between these proposals lies in different assumptions aboutthe way that non-relational and relational information are combined during early formulation, muchthe same way that production models differ in the extent to which they give either words or structurespriority during grammatical encoding.

A.E. Konopka, A.S. Meyer / Cognitive Psychology 73 (2014) 1–40 3

Addressing this debate, the two experiments reported in this paper tested whether the productionsystem supports flexibility in message and sentence formulation, allowing speakers to prioritizeencoding of either non-relational or relational information in different contexts. We first describethe key assumptions of each account of incremental sentence formulation. Then, we examine whetherchanges in the ease of encoding lexical and structural information favor one form of incrementalityover another during production of sentences like The dog is chasing the mailman. In Section 4, we out-line how and why speakers may flexibly shift between different planning strategies.

1.1. Incrementality in message and sentence formulation

Incrementality is often described as an adaptive property of the production system (Ferreira &Swets, 2002; Konopka, 2012; Levelt, 1989; Wagner, Jescheniak, & Schriefers, 2010): by allowing speak-ers to prepare a sequence of small units, incrementality distributes the cognitive burdens of planningover a series of words and thus lightens the load on the processor throughout the formulation process.For theoretical and practical reasons, most accounts of incrementality focus on encoding of the firstincrement of a message and sentence – i.e., on the selection of a starting point that initiates the map-ping of a preverbal message onto language (Bock et al., 2003, 2004). Experimental evidence for theseaccounts comes largely from eye-tracking studies where speakers are asked to describe pictures on acomputer screen: the pattern of speakers’ eye movements directed to different parts of the displayreveals what they encode with priority at various points in time when preparing their utterances, pro-viding insight into the online formulation of preverbal messages and their corresponding linguisticdescriptions (e.g., Brown-Schmidt & Tanenhaus, 2006; Griffin, 2004; Griffin & Bock, 2000; Meyer &Lethaus, 2004).

In the case of sentences describing coherent pictured events (e.g., a dog chasing a mailman), start-ing points may be selected on the basis of different types of information. One way of selecting a start-ing point is to quickly identify a salient character and encode it as the first character of the sentence(i.e., the grammatical subject in English). The strongest support for this proposal comes from studiesshowing that the perceptual salience of a character is sufficient to anchor this character as a startingpoint. For example, Gleitman et al. (2007) presented subliminal cues in the location of a characterin an upcoming pictured event and showed increased production of descriptions with the cued char-acter in subject position. Importantly, the cues directed speakers’ gaze to this character within 200 msof picture onset and speakers continued fixating this character preferentially until speech onset(approximately 2 s later).

This pattern of eye movements supports two conclusions: it suggests that the timing of visualuptake of information in a scene can control the order of encoding operations (the character that isfixated first is also encoded first) and that the first-fixated character can be assigned to subject posi-tion without extensive encoding of the second character. As a result, the first increment of the messageand sentence may consist only of information specific to one character (non-relational information)and not of information about its role in the event (relational information). The rest of the messageand sentence is then built to accommodate production of the cued character in subject position:speakers shift their attention and gaze to the second character only around speech onset and beginadding it to their message and sentence as the sentence object. This planning strategy is referred toas linear incrementality as it assumes that there is a simple, one-to-one mapping between visionand language: messages and sentences can be encoded, roughly, concept by concept and word byword (also see Osgood & Bock, 1977; Tomlin, 1995, 1997), and the relational structure of the eventemerges from the series of binding operations performed with the addition of each new conceptand new word to the sentence.

On a competing, hierarchically incremental account, the mapping of visual information onto lan-guage is mediated by formulation of a complex, higher-level message ‘‘plan.’’ Hierarchical incremen-tality predicts that relational processing initiates, rather than follows, the encoding of any oneincrement (Bock et al., 2003, 2004; Kuchinsky & Bock, 2010; Kuchinsky, Bock, & Irwin, 2011).Griffin and Bock (2000) provide support for this view by showing that, when characters in an eventdo not differ in perceptual salience, speakers have no clear preference for either character in the first400 ms of picture inspection. Convergence of fixations to the two characters in this time window is


interpreted as indicating a period of event apprehension where speakers encode the gist of the eventrather than favoring encoding of a single character (cf. Gleitman et al., 2007). In transitive events (e.g.,a dog chasing a mailman), apprehension involves the generation of a rudimentary message frameworkthat captures the who-did-what-to-whom causal structure of the event (one character chasinganother) and that identifies the two characters by the roles they play in the event (the chaser andthe chasee). This framework provides a form of top-down guidance at the outset of formulation: itallows speakers to select a starting point based on their construal of what the event is ‘‘about’’ andon their choice to take either character’s perspective instead of automatically assigning a salient char-acter to subject position without encoding its role in the event (analogous effects are found in thevisual search literature where cognitive relevance appears to quickly take precedence over perceptualsalience in controlling visual search patterns; see e.g., Henderson, Malcolm, & Schandl, 2009). Themessage framework also provides a blueprint for subsequent linguistic encoding: it controls deploy-ment of gaze at approximately 400 ms to the character selected to be the sentence subject and then,around speech onset, to the character selected to be the sentence object.

By defining the roles of the event characters on the basis of relational information shortly after pic-ture onset, hierarchical incrementality implies that early planning must be fairly extensive: incre-ments of the sentence generated before speech onset (e.g., The dog . . . in an active sentence) mustbe larger than later increments (. . . the mailman) as they include conceptual information about theevent as a whole and then linguistic information about the subject character. By comparison, planningof the second increment (the object character) is relatively fast: having already encoded thischaracter’s role in the event during the apprehension period, speakers must only complete linguisticencoding of its name to add it to the sentence once they have finished encoding the first character. Asimilar formulation process has been shown to support production of much simpler and overlearnedutterances like time expressions (e.g., eight twenty; Bock et al., 2003; Kuchinsky et al., 2011), wherepreparation times for the first element of the utterance (eight. . .) are longer that for the secondelement (. . .twenty).

In short, the two leading accounts of incrementality emphasize different criteria for the selection ofstarting points and make different assumptions about when and how speakers encode non-relationaland relational information within one utterance. Differences between these accounts are remarkablebecause they touch on fundamental questions about the way speakers formulate ‘‘thoughts’’ and theway that this information undergoes linearization: reliance on either non-relational or relational pro-cesses to initiate formulation has implications for the size and content of the first increment as well asfor planning of all subsequent increments (see Bock et al., 2004, for a review of a discussion that datesback to Wundt and Paul). At the same time, both accounts are intuitively appealing as speakers canplausibly employ either planning strategy to produce well-formed sentences: on the one hand, speak-ers can build sentences to talk about things that capture their attention in a bottom-up fashion(Gleitman et al., 2007) and, on the other hand, they can build sentences to express ideas that are orga-nized around some propositional content (Bock et al., 2004).

We outline a proposal for finding a middle ground in this debate: a continuum of incrementalitywith flexible selection of either planning strategy. Our approach largely follows from two recentfindings in the literature on incrementality and planning scope. First, different messages may lendthemselves to different types of planning strategies. Kuchinsky and Bock (2010) noted that the resultsoutlined above in support of linear and hierarchical incrementality were obtained in studies employ-ing pictures of events that differed in the ease of apprehension (i.e., the ease of relational encoding orthe ease of encoding event gist): unambiguous events in Griffin and Bock (2000) and substantiallymore difficult events in Gleitman et al. (2007). The unambiguous events elicited similar descriptionsacross speakers, suggesting high consensus in speakers’ interpretation of the events and thus of theunderlying message representations, while the ambiguous events elicited a wider range of descrip-tions, suggesting large differences in the content of speakers’ messages. Kuchinsky and Bock (2010)hypothesized that the harder it is to understand the gist of an event, the more likely speakers mightbe to use a linearly incremental strategy. They tested this hypothesis using a subliminal cuing para-digm similar to Gleitman et al. (2007): speakers described pictures of ‘‘easy’’ and ‘‘hard’’ events inwhich one of the characters had been cued before picture onset. Ease of conceptualization wasconfirmed with a codability measure that takes into account the number of different verbs used to


describe a depicted action (Shannon’s entropy): events that are consistently described with a small setof verbs are considered to be more codable, and thus permit faster encoding of event gist, than eventsdescribed with a wider range of verbs. Indeed, the perceptual salience of individual characters inKuchinsky and Bock’s (2010) study did not uniformly predict their assignment to subject position:the effect of cuing on selection of starting points was weaker in higher-codability than lower-codabil-ity events. This suggests that speakers did not rely on the salience of individual characters to select astarting point when they found it easy to do so on conceptual grounds. The results thus indicate adeparture from linearly incremental planning (as advocated by Gleitman et al., 2007) when relationalencoding is facilitated by the nature of the message (also see Myachykov, Garrod, & Scheepers, 2012,for an integrative approach to comparing linguistic and non-linguistic determinants of structurechoice).

Second, variability in planning scope can also result from a range of processing constraints. Forexample, differences in planning scope are often observed across studies eliciting sentences with sim-pler conceptual structures (e.g., The axe and the cup. . . or The axe is next to the cup). Relationshipsbetween objects in such sentences are arbitrary, which should generally favor linear (i.e., sequential)encoding. However, planning scope has been shown to range from one to two objects (see Konopka,2012, for a review). This variability can be attributed to several factors: it may reflect different goalsthat speakers have as they prepare their utterances (speed vs. fluency; e.g., Ferreira & Swets, 2002) andit may follow from language-specific and language-general processing bottlenecks. For example, theorder of encoding operations can be influenced by the phrasal syntax of a language (Brown-Schmidt & Konopka, 2008) and parallel processing can depend on the availability of processingresources (Konopka, 2012; Wagner et al., 2010).

When applied to production of sentences with more complex conceptual structures (like transitivesentences), these results imply that the timecourse of sentence formulation may vary systematicallybetween as well as within messages. Production may thus be neither strictly linearly incremental norstrictly hierarchically incremental. Instead, if the way that speakers assemble different pieces of infor-mation to produce full sentences can be controlled by a number of factors, the formulation processmay resemble linear, word-by-word planning and hierarchical, conceptually-driven planning in differ-ent contexts. More generally, the strong versions of linear and hierarchical incrementality may be bet-ter viewed as opposite ends of one continuum rather than representing two discrete planningstrategies.

What determines whether formulation of any one message and sentence falls towards one end orthe other end of this continuum? The hypothesis evaluated in this paper is that reliance on these plan-ning strategies should depend on the ease of non-relational and relational encoding, as well as oninteractions between these processes. Linear incrementality defines increments in terms of non-relational, character-specific information, whereas hierarchical incrementality gives precedence torelational over non-relational encoding. Thus if increments generated by applying a linearly incremen-tal or a hierarchically incremental planning strategy are encoded by prioritizing different types ofinformation, then differences in sentence formulation should be observed under two conditions. First,the timecourse of formulation should vary systematically across events with different non-relationaland relational properties (such as the ease of encoding individual characters and the ease of encodingevent gist). Second, formulation should shift from one end of the continuum to the other end of thecontinuum whenever processes responsible for encoding non-relational and relational informationbecome easier or harder to execute.

We report the results of two eye-tracking experiments that examined differences in the timecourseof formulation for descriptions of transitive events. In both experiments, participants saw anddescribed a list of pictures while their gaze and speech were recorded. The agent and patient charac-ters in the target events (n = 30 in each experiment) varied in ease of naming (character codability) andperformed actions that were easier or harder to describe (event codability; Kuchinsky & Bock, 2010). Inaddition, the ease of retrieving character names was manipulated in Experiment 1 via lexical priming,and the ease of generating active and passive structures was manipulated in Experiment 2 via struc-tural priming. Of these four variables, two provided a measure of the ease of non-relational encoding(character codability and ease of lexical retrieval) and two were more closely tied to relational encod-ing (event codability and the ease of assembling syntactic structures). Within each variable type, one


reflected item-specific properties and one was experimentally manipulated. Together, these variablescapture variability in encoding that can arise at the message level as well as the sentence level. Earlierwork showed that all four variables can influence sentence form (Bock, 1986a, 1986b; Kuchinsky &Bock, 2010), and detailed predictions with regard to the timecourse of formulation are listed below(Sections 1.2 and 1.3).

1.2. Encoding characters and actions

Message-level properties of events that bear on the formulation process are the identities ofindividual characters (non-relational information) and their relationship to one another (relationalinformation). The pictured events used in Experiments 1 and 2 varied on both dimensions. UsingKuchinsky and Bock’s (2010) approach, estimates of the ease of encoding characters and actions werebased on the heterogeneity of speakers’ descriptions.

Variability in event descriptions is expected in open-ended production tasks because differentspeakers can interpret the same event in different ways. For example, speakers can choose to empha-size different aspects of a character’s identity (e.g., man vs. policeman) or take different perspectives onthe same action (e.g., kicking vs. pushing). For character naming, the index of conceptual difficulty isthus heterogeneity in speakers’ noun choice: characters referred to consistently with a small set ofnouns are assumed to be more codable than characters with lower name agreement. For actions,the index of conceptual difficulty is heterogeneity in verb choice: events that are consistentlydescribed with a small set of verbs are assumed to be more codable than events eliciting a wider rangeof verbs.1

We first examined whether character codability and event codability influenced what speakers saidand then whether they influenced how speakers assembled their sentences. If formulation is flexible,then variations in character codability and event codability across items should shift control of formu-lation from a relational to a non-relational source, and vice versa. Effects of these variables may beobserved at two points in the formulation process: first, during selection of a starting point and encod-ing of the first character (Gleitman et al., 2007, vs. Kuchinsky & Bock, 2010), and, second, during theaddition of the second character to the sentence. Since the two experiments used a highly overlappingset of target items, the same predictions apply to both experiments.

First, variations in character codability should produce accessibility effects in sentence form and inearly gaze patterns to target characters. Speakers have a strong preference to begin sentences withaccessible characters (Altmann & Kemper, 2006; Bock, 1987b; Bock & Irwin, 1980; Bock & Warren,1985; Branigan, Pickering, & Tanaka, 2008; Christianson & Ferreira, 2005; Ferreira, 1994; McDonald,Bock, & Kelly, 1993; Prat-Sala & Branigan, 2000), so easy-to-name characters should become subjectsmore often than harder-to-name characters. These effects are generally compatible with a strong, lin-early incremental account of planning where starting points are selected based on the ease of encod-ing non-relational information. Consequently, we expected character codability to also predictassignment of first-fixated characters to subject position: first-fixated characters should become sub-jects more often when they are easy to name than when they are harder to name, demonstrating adirect link between character accessibility and selection of starting points. The strong version of linearincrementality also predicts that this selection process should occur during the earliest stages of for-mulation: fixations to the two characters should diverge quickly after picture onset (within 200 ms ofpicture onset; Gleitman et al., 2007) when an ‘‘easy’’ character in produced in subject position andspeakers should continue fixating the subject character preferentially until speech onset.

1 In principle, heterogeneity in noun and verb choice can arise both during message-level encoding as well as duringlexicalization (sentence-level encoding). For example, after identifying a character, speakers may need to select one noun from arange of synonymous nouns to express this information linguistically. Similarly, having apprehended what action two charactersare performing, speakers may need to select one verb from a range of nearly synonymous verbs. Fine-grained semantic differencesbetween nouns and verbs that different speakers use in their descriptions may thus reflect either fine-grained differences ininterpretation, fine-grained differences in the accessibility of individual nouns and verbs, or both. For present purposes, charactercodability and event codability are used as an umbrella measure of how easily speakers identify individual characters (non-relational information) and express what an event is ‘‘about’’ (relational information), regardless of the locus or origin ofdifferences in word choice.


In contrast, the effect of event codability on early formulation should be to reduce the impact of firstfixations and character codability on selection of starting points. Replicating Kuchinsky and Bock(2010), speakers should begin their sentences with first-fixated characters or easy-to-name charactersless often in higher-codability than lower-codability events. Early eye movements should also showsensitivity to higher-level event properties (see Bock et al., 2003; Dobel, Gumnior, Bolte, &Zwitserlood, 2007; Hafri, Papafragou, & Trueswell, 2012, for demonstrations of rapid encoding of eventgist). Speakers should be less likely to prioritize encoding of one character over the other character inthe first 400 ms of picture inspection in higher-codability events than in lower-codability events;instead, they should direct their gaze preferentially to the subject character later in higher-codabilityevents, resulting in slower divergence of fixations to the two characters in higher- than lower-codabil-ity events immediately after picture onset. In other words, formulation should begin with a periodwhere speakers distribute their attention roughly evenly between the two characters when the gistof an event is easy to encode, as predicted by the strong version of hierarchical incrementality(Bock et al., 2003, 2004; Griffin & Bock, 2000).

Second, we extend the predictions of linear and hierarchical incrementality to processes requiredto add the second character to the developing sentence. We propose that differences in planningstrategies across events may also be observable in the timing of gaze shifts from the first characterto the second character around speech onset. The duration of gazes to a character immediatelybefore production of its name is assumed to index the speed of lexical retrieval (name-related gazes;Griffin, 2004; Meyer & Lethaus, 2004), so, in all events, speakers were expected to fixate easy-to-name (high-codability) subject characters for less time than harder-to-name (lower-codability)subject characters. However, the extent to which speakers encoded relational information aboutthe event (i.e., information about both the first and second character) at the outset of formulationshould also influence the length of gazes to the subject character. Linear incrementality predicts thatearly planning does not extend beyond the first character, so this planning strategy should confer nobenefits for encoding of the second character mid-way through the sentence: the timing of gazeshifts away from the first character and towards the second character should depend on charactercodability but should not differ between higher-codability and lower-codability events. In contrast,hierarchical incrementality assumes early encoding of relational information after picture onsetand thus allows for top-down control of linguistic encoding: the second character should be easierto add to the developing sentence when formulation is supported by a conceptual framework, sospeakers may initate gaze shifts to the second character earlier in higher-codability than lower-codability events.

Finally, if changes in formulation are observed across events differing in codability of charactersand actions, interactions between character codability and event codability should highlight the rela-tive sensitivity of the production system to non-relational and relational variables. On the one hand, ifthe production system generally favors linearly incremental planning, the ease of encoding non-relational information should control the extent to which relational properties of an event also influ-ence formulation: character codability should be the strongest predictor of formulation, allowingeffects of event codability to be observable relatively late in the formulation process or only whencharacters are difficult to encode (i.e., in events where speakers may need to fall back on other typesof information to formulate their sentences). On the other hand, if the production system generallyfavors hierarchically incremental planning, event codability should control the extent to which prop-erties of agents and patients also influence formulation: character codability should play a weak rolein formulation for high-codability events and a stronger role for low-codability events (i.e., eventswhere speakers may need to fall back on non-relational information to formulate their sentences).

1.3. Retrieving words and assembling structures

Message-level planning involves encoding of non-relational and relational information, so themapping of conceptual representations onto language also takes place via processes responsible forencoding two types of information. Production models agree that the mapping from concepts tolanguage occurs via the process of lemma selection (for content words) and function assignment(for links between concepts, thematic roles, and structure-building procedures). In this sense, words


and structures carry non-relational and relational information respectively: nouns (like dog and mail-man) describe referents, rather than the relationship between referents, whereas structures (likeactives and passives) express these relationships, irrespective of the identity of the referents.2

Since speakers are able to pass on conceptual information to linguistic encoding as soon as itbecomes available, both lexical and structural processes can be expected to systematically influencethe timecourse of formulation. Importantly, lexical and structural processes have different implica-tions for incrementality because words and structures have an inherently different scope (Konopka,2012): retrieving a character name does not require planning of material beyond this character,whereas generation of a structure inherently involves preparation of information that extends beyonda single character. Thus prioritizing lexical encoding is generally compatible with the main tenets oflinear incrementality, and prioritizing encoding of structural information is more compatible withhierarchical incrementality.

As outlined earlier, discussions about the role of words and structures in formulation bear a strongsimilarity to questions about lexical–structural integration – i.e., the long-standing debate about lex-ical and structural guidance in grammatical encoding (Bock, 1987a; see Pickering & Ferreira, 2008, fora review). Sentence form is a product of both lexical and structural constraints, but lexicalist andabstract structural accounts assume that speakers prioritize encoding of either words or structures:in a lexical system, words trigger structural assembly (lexical guidance), and in an abstract structuralsystem, structures can be generated without lexical support (structural guidance). The types of depen-dencies between words and structures described by these accounts have the same implications forformulation as linear and hierarchical incrementality: lexical guidance assumes that non-relationalprocesses take precedence over relational processes, while structural guidance gives abstractstructures a more prominent role in shaping sentence form.

One approach to testing for effects of lexical and structural guidance on formulation is to experi-mentally vary the ease of lexical and structural encoding. In the current experiments, we manipulatethese processes via lexical and structural priming. Lexical priming involves presenting speakers withwords that are semantically or associatively related to a referent in the target picture (e.g., pony or milkbefore a picture of a horse kicking a cow; Bock, 1986b). Processing of the prime words increases theactivation and hence the accessibility of target words, and thus increases production of sentences withprimed, easy-to-name characters in subject position. Similarly, structural priming involves exposingspeakers to syntactic structures that may be used to describe target events, and thus increases thelikelihood of speakers using the primed structure on the target trial (Bock, 1986a).

Experiment 1 used lexical primes embedded in intransitive sentences to increase the accessibilityof the agent and patient characters in target events, and Experiment 2 used structural primes to facil-itate assembly of a transitive structural frame. The paradigms were adapted from Bock (1986a,1986b): on prime trials, speakers saw pictured events and heard recorded descriptions, while on tar-get trials, they were asked to describe new pictures themselves. Speakers were expected to producemore active sentences after agent-related primes than patient-related primes (Experiment 1) andmore active sentences after active primes than passive primes (Experiment 2). Importantly, if formu-lation is flexible, then any changes in structure choice resulting from lexical and structural primingshould also be accompanied by changes in the timecourse of formulation.

Specifically, facilitating encoding of individual characters in Experiment 1 should result in an earlyaccessibility effect: a first-fixated character should be produced in subject position more often if it hadbeen primed than if it had not been primed, and speakers should spend more time fixating primedsubject characters than unprimed subject characters immediately after picture onset (0–400 ms). Bothresults would indicate priority encoding of accessible characters before less accessible characters. Thisis analogous to the predicted effect of character codability on early formulation (Section 1.2), and

2 This discussion of content words is limited to nouns for maximum contrast between linguistic elements that express non-relational and relational information. Naturally, relationships between referents are also expressed lexically with content wordslike verbs. Verb choice can provide an indication of speakers’ interpretation of the conceptual structure of the event, and indeed,the measure of event codability used to assess the ease of relational encoding is based on verb heterogeneity. Thus for presentpurposes, verbs can be viewed as being more similar to structures in terms of their relational content than nouns (see Fisher,Gleitman, & Gleitman, 1991, for a detailed discussion of the semantics of different verb classes).


indicates a shift towards linearly incremental planning. In contrast, facilitating encoding of sentencestructure in Experiment 2 should reduce the likelihood of speakers fixating one character preferen-tially over the other character immediately after picture onset: speakers should be more likely to dis-tribute their attention between two characters when producing a primed structure than an unprimedstructure. This is similar to the predicted effect of event codability on formulation (Section 1.2) andillustrates a shift towards hierarchical incrementality.

Later in the formulation process (i.e., between 400 ms and speech onset), the lexical and structuralprimes should both also influence the timing of gaze shifts from the first to the second character: lex-ical primes should reduce the length of gazes on a primed subject character by facilitating encoding ofits name and structural primes should reduce the length of gazes on the subject character by facilitat-ing encoding of the entire event. Importantly, despite similar outcomes, the reasons for these effectscan be traced back to qualitative differences in planning strategies in the two experiments.

1.4. Overview of experiments

In sum, in two experiments, we undertook a systematic analysis of the influence of non-relationaland relational variables on the timecourse of formulation for simple event descriptions. Similar resultswere expected for the two variables influencing the ease of non-relational processing (character coda-bility and lexical accessibility) and the two variables influencing the ease of relational processing(event codability and ease of generating linguistic structures).

Analyses in each experiment first verified whether all variables had the expected effect on speak-ers’ descriptions of target events (i.e., structure choice). First, character codability was expected toinfluence the assignment of characters to subject or object position based on their relative ease ofnaming in both experiments, and lexical priming was expected to produce a similar effect in Experi-ment 1. Second, structure choice was evaluated with respect to event codability and structural prim-ing. Event codability was not expected to influence structure selection on its own as the differencebetween an active frame and a passive frame is not inherently linked to the ease of encoding eventgist, but the structural primes in Experiment 2 were expected to produce the well-documented struc-tural priming effect.

After confirming effects of these variables on structure selection, we examined whether and howthey also shaped the timecourse of formulation in active sentences (i.e., descriptions of events withthe preferred active structure; see Van de Velde, Meyer, & Konopka, 2014, for discussion of formula-tion of sentences with the dispreferred passive structure). We began by testing whether first fixationspredicted sentence form across items and conditions. Timecourse analyses were then carried out tocompare the distribution of fixations to the two characters over time in early (0–400 ms) and late(400 ms – speech onset) time windows across items and conditions. To summarize the predictions,character codability and lexical priming were expected to (a) favor selection of the first-fixated char-acter as the starting point and (b) favor priority encoding of this character after picture onset (thestrong version of linear incrementality). In contrast, event codability and structural priming wereexpected to (a) reduce the impact of first fixations on selection of starting points, (b) favor priorityencoding of relational information about the event after picture onset, and (c) influence the timingof gaze shifts from the first character to the second character around speech onset (the strong versionof hierarchical incrementality). We highlight effects consistent with linear and hierarchical incremen-tality throughout the results sections, and we refer to effects that are consistent with both accounts assupporting weaker versions of linear and hierarchical incrementality.

2. Experiment 1

Eye-tracked participants described a long series of pictures, including 30 target pictures of two-character events. They were asked to mention all characters shown in each picture, but, to approxi-mate production in more naturalistic situations, they received no further instructions about sentencecontent or form. Event and character codability were estimated post hoc for each target picture. Coda-bility ratings for events and agents in Experiments 1 and 2 were highly correlated (both rs > .87),


showing high stability in the types of descriptions speakers produced to describe the events andwarranting a direct comparison of results across experiments.

The ease of character naming in target pictures in Experiment 1 was additionally manipulated withlexical priming. Target pictures were preceded by primes where speakers saw a picture of an intran-sitive event and heard a recorded intransitive description. The referent shown in the prime pictureswas semantically or associatively related to the agent or to the patient in the following target picture(the agent and patient prime conditions respectively) or to neither character (the neutral prime condi-tion). The agent and patient characters were thus either primed or unprimed. The neutral conditionserved as a baseline to assess the overall likelihood of speakers using active and passive syntax todescribe the target transitive events.

Timecourse analyses assessed differences and changes in the formulation of active descriptions forthe different types of events and after the three types of primes. On the hypothesis that the ease ofcharacter naming determines the extent to which speakers prioritize encoding of a single characterat the outset of formulation, speakers should be more likely to engage in linearly incremental thanhierarchically incremental planning when preparing sentences that begin with an accessible character(a highly-codable character or a primed character); event codability should have the opposite effect onformulation.

2.1. Method

2.1.1. ParticipantsFifty-four native speakers of Dutch (mostly university students; 48 female) from the Nijmegen area

participated for payment. Four participants were replaced because they produced very few scorableresponses on target trials.

2.1.2. Materials and designThere were four types of trials: target trials, prime trials, filler trials, and word trials. On target trials,

speakers saw pictures of transitive events (see Appendix A; pictures were adapted from Bock, 1986b,and from images available in the Microsoft clipart database). There were 20 items with animate agents(13 items with human agent and 7 with animal agents), and 10 with inanimate agents. To increaseproduction of passive sentences, 23 items had animate patients (20 items had human patients, 3had animal patients) and 7 had inanimate patients.3

Pictures shown on prime trials were one-character events. They were accompanied by a recordedintransitive description produced by a native Dutch speaker. The characters named in these sentenceswere semantically related to the agent (e.g., wolf), the patient (e.g., salesman), or to neither character(e.g., umbrella) in the following target picture (in this case, a dog chasing a mailman). Semantic relat-edness was verified with LSA norms (Latent Semantic Analysis; http://lsa.colorado.edu): across allevents, agent primes had a stronger relationship to agents than patients (.37 vs. .09; t(28) = 6.20),and patient primes had a stronger relationship to patients than agents (.23 vs. .12; t(29) = 3.06). Neu-tral primes were not related to either character (.05 and .08 for the relationship to the agent andpatient respectively).

The remaining trials were unrelated to the prime and target pictures. On filler trials (n = 103),speakers saw pictures that could be described with a variety of structures (e.g., intransitive, dative,reflexive sentences). On 90 filler trials, speakers produced a description, and on 13 trials, they sawa picture and heard a recorded description.

Word trials were introduced to present the experiment as a memory task to participants (see Bock,1986a, 1986b). On these trials (n = 25), participants saw a printed word appearing in the center of thescreen: 15 of these words had been used previously in the experiment (e.g., they were names of char-acters shown in earlier pictures) and 10 were new.

3 As expected, speakers produced more active sentences to describe events with animate agents than inanimate agents, andmore passive sentences to describe events with animate patients than inanimate patients in both experiments. Since the effects ofanimacy on sentence form are well-documented in the literature, we omit discussion of animacy in this paper and focus only oncharacter codability.

http://lsa.colorado.edu


The design of the experiment included one three-level factor (Prime condition: agent primes,patient primes, neutral primes). Two mirror-reversed versions of each target picture were created,one in which the agent appeared on the left hand-side and one in which the agent appeared on theright hand-side of the screen. Crossing this factor with the priming manipulation produced six differ-ent lists of stimuli, with each target picture being presented in a different Prime condition and with adifferent spatial layout of characters on each list. All analyses collapsed across the two mirror-reversedversions of each item. Within lists, there were 10 target pictures in each Prime condition, and no twoprime-target pairs from the same condition were presented in succession. The prime-target pairs wereseparated by 4–5 intervening unrelated trials (filler trials and word trials).

2.1.3. ProcedureParticipants were seated at an Eye-link 1000 eye-tracker (500 Hz sampling rate). Instructions

appeared on the screen and were paraphrased by the experimenter. Participants were told that theywould see a series of pictures and of single words. Each trial began with a fixation point at the top ofthe screen: participants had to look at the fixation point and press a key to continue. On picture trials,they then saw a picture of an event and their task was to describe the event with one sentence, men-tioning all characters shown in the event. On a subset of picture trials, participants first saw the wordLISTEN printed in the center of the screen and then a pictured event accompanied by a recorded sen-tence: here, their task was to listen to the sentence and then repeat it out loud. On word trials, partic-ipants saw a printed word in the center of the screen instead of a picture: they were instructed to readthis word out loud and decide whether they had produced it before in the experiment by saying ‘‘yes’’or ‘‘no.’’

2.1.4. Sentence scoringSentences produced on target trials were scored as actives, full passives, or sentences with other

constructions. The latter category included truncated passives, get-passives, intransitive sentences,sentences beginning with a ‘‘There is/are. . .’’ construction, and sentences with indefinite pronouns(e.g., someone). Two items were excluded from the analyses as they elicited a very small number oftransitive responses and one item was excluded for technical reasons.

In the remaining dataset, trials were excluded if the first fixation in that trial was not on thefixation point at the top of the screen (80 trials) or if onsets were longer than 5 s and longer than 3standard deviations from the grand mean (11 sentences). The final dataset consisted of 648 sentences(.71 actives, .29 full passives).

Sentences containing disfluencies were included in the analyses of structure choice and in time-course analyses because disfluencies are a normal occurrence in unprepared speech. Sentences withdisfluencies before the first article or first noun were, however, not included in analyses of speechonset, leaving 627 fluent sentences.

2.1.5. Character codability and Event codability scoringCharacter codability and Event codability were estimated with Shannon’s entropy based on the dis-

tribution of responses included in the analyses (see Kuchinsky, 2009).4 All the different referentialterms that speakers used in their descriptions were included in the codability estimates. Higher codabil-ity scores for agents and patients indicate lower heterogeneity in speakers’ choice of referential termsand thus greater ease of identification and naming. Similarly, higher codability scores for events indicatelower heterogeneity in speakers’ descriptions of the action shown in the event, and thus greater ease ofevent apprehension and gist extraction.

4 Shannon’s entropy (H) = �P

pi log2(pi). Pi is the probability of the ith word. For example, the event showing a fireman saving aboy was consistently described with the verb save (low verb heterogeneity and thus high event codability; H = 0); the eventshowing a shark attacking a man was described with verbs like attack, bite, eat, and surprise (high verb heterogeneity and thuslower event codability; H = 1.56). Similarly, the agent and patient in the event showing a horse kicking a cow were consistentlydescribed with the nouns horse and cow (high agent and patient codability; H = 0); characters in the event showing a bee stinging aman were described with nouns like bee, wasp, mosquito, insect, and creature for the agent (lower agent codability; H = 2.32), andnouns like man, boy, and person for the patient (lower patient codability; H = 1.13).


As expected, the codability scores showed large between-item differences (see Table 1 for meanscores after median splits), allowing analyses of the effects of these variables on structure choiceand formulation across a range of events. Event codability scores were not correlated with Agent orPatient codability (r = .17 and �.31, ns., respectively), so the identity of the characters had little bear-ing on the ease of comprehending the events. Agent and Patient codability were, however, positivelycorrelated (r = .42, p < .05): items with easier-to-name agents contained easier-to-name patients.Patient codability scores were thus residualized on Agent codability for analyses of sentence form;since properties of the patients did not reliably predict sentence form, this factor was then droppedfrom all analyses. Importantly, codability ratings for agents and patients did not differ across Primeconditions (all ps > .3), showing that the lexical primes did not influence speakers’ choice of referentialterms for these characters and thus did not contribute further to variability in naming.

2.1.6. AnalysesAnalyses of structure choice and speech onsets were conducted with mixed logit models and linear

mixed effects models respectively in R (Baayen, Davidson, & Bates, 2008; Jaeger, 2008). The modelsincluded a combination of Event codability, Agent codability (continuous predictions), the locationof First fixations, and Prime condition (categorical predictions) as listed below. All predictors werecentered. For clarity, the effects of Event and Agent codability are shown in all figures following amedian split into higher- and lower-codability events (‘‘easy’’ and ‘‘hard’’ events) with higher- andlower-codability agents (‘‘easy’’ and ‘‘hard’’ agents). Performance in the three Prime conditions wascompared with two orthogonal contrasts motivated by the data (as listed in all tables).

Analyses were carried out in four steps. The first analysis considered effects of First fixations on sen-tence form (Section 2.2.1), and the second analysis tested the influence of the lexical primes on sentenceform (Section 2.2.2). Both analyses also tested for interactions with Event and Agent codability. Thethird analysis tested whether speech onsets were sensitive to differences in ease of encoding acrossitems and conditions as well (Section 2.2.3). Finally, timecourse analyses of agent-directed fixationswere carried out for with quasi-logistic regressions for active sentences (Section 2.2.4; Barr, 2008).

In all cases, to arrive at the simplest best-fitting models, full models including all interactionsbetween factors were simplified to include only reliable interactions that improved model fit. Randomslopes for fixed factors were included where mentioned only if they improved model fit (models withthe full random structure often failed to converge; similar results were obtained in models with themost complex possible random structure and are therefore not reported; cf. Barr, Levy, Scheepers, &Tily, 2013). All effects were considered to be reliable at p < .05, unless indicated otherwise.

2.2. Results

2.2.1. First fixations and sentence formOn the majority of scored target trials, first fixations were directed to the agent (.65). Speakers also

directed more first fixations to the agent after agent primes (.66) than after neutral primes (.64) andpatient primes (.64), but differences between conditions did not reach significance.

Table 1Means (and standard deviations) of Event codability and Agent codability scores for higher- and lower-codability events withhigher- and lower-codability agents in Experiments 1 and 2.

Event codability Agent codability Event codability score Agent codability score

Experiment 1High High .85 (.58) .42 (.46)High Low .46 (.57) 1.70 (.59)Low High 1.84 (.32) .33 (.40)Low Low 2.11 (.42) 1.92 (.45)

Experiment 2High High .98 (.41) .41 (.32)High Low .66 (.49) 2.17 (.44)Low High 1.99 (.38) .32 (.28)Low Low 2.36 (.51) 2.09 (.69)


More importantly, first fixations predicted selection of starting points (Fig. 1a): speakers produced.12 more actives if they looked first at the agent than if they looked first at the patient (.75 vs. .63;b = .61, z = 2.09). Supporting linear incrementality, this result shows that selection of a starting pointcan be influenced by shifts of visual attention and thus by the timing of the uptake of visual informa-tion (Gleitman et al., 2007; Kuchinsky & Bock, 2010). There were no interactions with Prime conditionor with the two Codability predictors.

2.2.2. Lexical primes, codability, and sentence formLexical primes reliably influenced sentence form (Fig. 2a; Table 2): speakers produced fewer active

sentences after patient primes than after other primes (agent and neutral primes; the first contrast forPrime condition in Table 2). Production of active sentences after agent primes and after neutral primesdid not differ (the second contrast for Prime condition in Table 2). The asymmetry in priming effectsafter agent and patient primes shows that only priming of the patient character influenced speakers’selection of an active or passive structure.

Priming effects were additionally modulated by Agent codability and Event codability. Speakersproduced more active sentences beginning with ‘‘easy’’ agents than ‘‘hard’’ agents (.80 vs. .60).Importantly, the lexical primes influenced sentence form only in events with ‘‘harder’’ agents(Fig. 2b): speakers produced fewer actives after patient primes than other primes (agent and neutralprimes) to describe these events, while events with ‘‘easier’’ agents were less amenable to priming(resulting in an interaction between Prime condition and Agent codability; see the first contrast for

Fig. 1. Proportions of active sentences produced in (a) Experiments 1 and (b) Experiment 2 after agent-directed and patient-directed first fixations across conditions. Error bars are by-participant standard errors, (c) shows proportions of active sentencesproduced across condition for ‘‘easier’’ and ‘‘harder’’ events in Experiment 2.

Fig. 2. Proportions of active sentences produced in Experiments 1 and 2 (a) across all items, (b) with respect to Agent codabilityand (c) with respect to Event codability. Error bars are by-participant standard errors.


this interaction in Table 2). Thus the lexical primes were effective primarily in cases where speakersgenerally preferred to postpone encoding the agent (i.e., in events with ‘‘hard’’ agents).

A similar effect was observed with respect to Event codability (Fig. 2c). The lexical primes influ-enced sentence form primarily in ‘‘harder’’ events: again, speakers produced fewer active sentencesafter patient primes than after other primes (agent and neutral primes), while descriptions of ‘‘easier’’events were less more amenable to priming (resulting in an interaction between Event codability andPrime condition; see the first contrast for this interaction in Table 2). The direction of the effect isconsistent with Kuchinsky and Bock’s (2010) finding that perceptual cues have a stronger effect onselection of starting points in ‘‘hard’’ events: here, manipulating the ease of encoding patients withlinguistic cues (lexical primes) instead of non-linguistic cues influenced sentence form to a greaterextent in ‘‘hard’’ events, where starting points were difficult to select on conceptual grounds, thanin ‘‘easy’’ events, where starting points were easier to select on conceptual grounds.

Table 2Production of active sentences in Experiments 1 and 2, as a function of Prime condition, Agent codability, and Event codability. (p)and (i) include addition of random by-participant and by-item slopes.

Predictor Experiment 1 Experiment 2

Est. SE z-Value Est. SE z-Value

Intercept �1.89 .64 �2.93� �1.24 .38 �3.27�

Prime condition �.56 .31 �1.80� �.49 (p, i) .16 �3.09�

Exp. 1: patient prime vs. other primesExp. 2: passive prime vs. other primes

Prime condition �.04 .33 �.11 �.07 (p, i) .19 .41Exp. 1: agent prime vs. neutral primeExp. 2: active prime vs. neutral prime

Agent codability .23 .72 .33 .28 .37 .76Event codability �.63 .73 �.86 .01 .44 .01Prime condition � Agent codability �1.30 .41 �3.20� .14 .17 .85

Exp. 1: patient prime vs. other primesExp. 2: passive prime vs. other primes

Prime condition � Agent codability �.28 .42 �.66 .46 .20 2.30�

Exp. 1: agent prime vs. neutral primeExp. 2 active prime vs. neutral prime

Prime condition � Event codability �.88 .43 �2.07� – – –Exp. 1: patient prime vs. other primesExp. 2: passive prime vs. other primes

Prime condition � Event codability �.26 .48 �.55 – – –Exp. 1: agent prime vs. neutral primeExp. 2: active prime vs. neutral prime

� p < .05.� p < .10.


2.2.3. Speech onsetsActive and passive sentences had comparable onsets (1900 and 1859 ms respectively) and onsets

did not differ reliably across Prime conditions. Onsets varied only with the ease of naming the agent:sentences describing events with ‘‘easy’’ agents were initiated more quickly (1842 ms) than sentenceswith ‘‘harder’’ agents (1939 ms; b = .12, z = 2.09, for the main effect of Agent codability). There was nointeraction between Agent codability and Sentence form, suggesting that agents were encoded withpriority in both active and passive sentences and thus that speakers had a strong preference forplacing agents in subject position.

2.2.4. Timecourse analysesQuasi-logistic regressions (performed by participants and by items) compared the proportions of

agent-directed fixations across items and conditions for active sentences over time (Barr, 2008).5 Fix-ations were first binned into consecutive 10 ms time samples and then aggregated into 200 ms time bins.An empirical logit was calculated for each time bin indexing the log odds of speakers fixating the agent inthat time bin (out of the total number of fixations to the agent, patient, and to empty areas on the screenobserved in that time bin). Regressions were performed on the empirical logits.

We first tested the effect of event properties that were not manipulated experimentally by compar-ing the distribution of agent-directed fixations with respect to Agent codability and Event codability(Section 2.2.4.1). Codability scores were included as categorical predictors in the by-participant anal-yses (following a median split into higher- and lower-codability events and agents) and as continuouspredictors in the by-item analyses. Next, a separate analysis compared production of sentencesdescribing events with ‘‘easier’’ and ‘‘harder’’ agents across Prime conditions, as character codability

5 Performing separate by-participant and by-item analyses is a way of compensating for the non-independence of individualfixations. Briefly, fixations to characters in an event on any given trial by the same speaker are strongly intercorrelated.Aggregating data within a condition can reduce the degree to which non-independence inflates statistical outcomes. Aggregationcan be done across items and across participants, as is typically done to calculate F1 and F2 statistics in the ANOVA framework.


and character priming were expected to produce analogous effects (Section 2.2.4.2). There were nointeractions between Prime condition and Event codability, so this analysis is not reported.

Three time windows were chosen for examination in each analysis based on three theoreticallyimportant distinctions. The first time window included the period between 0 ms (picture onset)and 400 ms (i.e., two consecutive bins of 200 ms each): on Griffin and Bock’s (2000) account, speakersmay select a starting point in this time window on the basis of their construal of the gist of the event(hierarchical incrementality), or, on Gleitman et al.’s (2007) account, on the basis of non-relationalproperties of the two characters (linear incrementality).6 It was expected that formulation would bemore hierarchically incremental in high-codability events and more linearly incremental in events withhigh-codability agents. Priming character names in this experiment was also expected to result in a shifttowards linear incrementality.

The second and third time windows included the period between 400 ms and speech onset thatcorresponds to retrieval of the first character name (name-related gazes). This time window includesa segment with increasing fixations (400–1000 ms, i.e., three 200 ms time bins) and decreasing fixa-tions to this character (1000–1800 ms in Experiment 1, and thus four 200 ms time bins; 1000–2200 ms in Experiment 2, and thus six 200 ms time bins). The length of gazes on the agent and thusthe timing of gaze shifts from the agent to the patient were expected to reflect the ease of encoding theagent and to show when speakers were ready to begin adding the patient to the sentence.

The models included all predictors as additive factors and only interactions that contributed tomodel fit (p < .10) and that were reliable at pMCMC < .05 (for models without random slopes) orp < .05 (for models with random slopes), unless stated otherwise. The by-item analyses had less sta-tistical power, so interactions in these analyses that were reliable but did not improve model fit (rel-ative to models without these interactions) are reported for comparison with the by-participantanalyses. Main effects of a variable in these models indicate differences in fixations at the start of agiven window (i.e., the first time bin in a given window), and interactions with the Time variable (Timebin) indicate changes in the slopes of fixations over time (i.e., changes between the first time bin andsubsequent time bins in a given window).

2.2.4.1. Effects of Agent codability and Event codability. Fixations between 0 and 400 ms. Fig. 3a and bshows the timecourse of formulation for descriptions of ‘‘easy’’ and ‘‘hard’’ events with ‘‘easy’’ and‘‘hard’’ agents. Overall, speakers quickly directed their attention to the agent between 0 and 200 msof picture onset and then briefly looked back to the patient by 400 ms.

There were more fixations to ‘‘easy’’ agents than ‘‘hard’’ agents within 200 ms of picture onset (amain effect of Agent codability; Table 3a), and this difference increased in the 200–400 ms time bin(resulting in an interaction of Agent codability with Time bin). The preference for fixating ‘‘easy’’agents is consistent with linear incrementality as it shows immediate effects of character-specificproperties on early formulation.

Importantly, differences in the distribution of early fixations in this time window were alsomodulated by Event codability (resulting in an interaction between Event and Agent codability inthe by-participant analysis). The difference in fixations to ‘‘easy’’ and ‘‘hard’’ agents was larger inlower-codability events than in higher-codability events: speakers were less likely to fixate ‘‘easy’’agents in ‘‘easy’’ events than to fixate ‘‘easy’’ agents in ‘‘hard’’ events, but were more likely to fixate‘‘hard’’ agents in ‘‘easy’’ events than to fixate ‘‘hard’’ agents in ‘‘hard’’ events. This shows sensitivityto character properties when the event is hard to encode and less sensitivity to character propertieswhen the event is easy to encode, which is broadly consistent with hierarchical incrementality.

Fixations between 400 and 1000 ms. Having fixated agents with priority at the outset of formulation(0–400 ms), speakers did not continue formulating sentences with an easy-to-name agent in subjectposition. Instead, they shifted their gaze back to the patient by 400 ms, suggesting that they also pre-ferred to encode information about the second character relatively early in the formulation process.Fig. 3 shows that the shift of gaze away from the agent was larger in items with ‘‘easy’’ agents, so there

6 Quasi-logistic regressions only test for linear effects in the data. Aggregating the data into 200 ms bins averages over theinflection point observed at 200 ms (the rise in fixations observed until 200 ms and the fall in fixations observed before 400 ms),and thus does not require adding a quadratic term to the model.

Fig. 3. Timecourse of formulation for active sentences in Experiment 1 and 2 with respect to Event codability (‘‘easy’’ vs. ‘‘hard’’events) and Agent codability (‘‘easy’’ vs. ‘‘hard’’ agents).


were fewer fixations to ‘‘easy’’ agents than ‘‘hard’’ agents at the start of the 400–1000 ms time window(i.e., at 400–600 ms; a main effect of Agent codability; Table 3b). In contrast, when speakers fixatedagents to a lesser extent before 400 ms (‘‘hard’’ agents), they were more likely to immediately turntheir gaze to the agent after 400 ms.

Table 3Results of regressions comparing fixations to the agent in active sentences describing ‘‘easier’’ and ‘‘harder’’ events with ‘‘easier’’and ‘‘harder’’ agents in Experiment 1. (s) indicates the inclusion of random slopes, (a) indicates that the effect does not reliablyimprove model fit.

Effect By-participants By-items

Est. SE t-Value Est. SE t-Value

(a) 0–400 msIntercept �1.70 .05 �34.93� �2.02 .02 �88.49�

Time bin 9.16 (s) .25 37.22� 10.34 (s) .14 72.50�

Event codability �.41 .07 �6.06� �.12 .03 �4.30�

Agent codability �.26 .07 �3.52� �.14 .03 �4.77�

Time bin � Event codability 2.20 .41 5.37� 1.20 .18 6.72� (a)Time bin � Agent codability �2.98 .44 �6.74� �1.93 .18 �10.82� (a)Event codability � Agent codability �.62 .10 �6.44� .05 .04 1.23

(b) 400–1000 msIntercept �.14 .02 �6.08� �.26 .02 �13.65�

Time bin 1.66 .09 18.54� 2.79 .05 55.23�

Event codability .41 .05 8.82� .24 .03 9.61�

Agent codability .25 .04 7.05� .28 .02 14.07�

Time bin � Event codability �1.49 .18 �8.32� �.78 .07 �11.09� (a)

(c) 1000–1800 msIntercept .55 .03 20.37� .90 .02 57.07�

Time bin �1.15 .06 �20.22� �1.61 (s) .04 �44.47�


Agent codability .05 .04 1.42 .36 .01 26.13�

Event codability � Agent codability �.24 .07 �3.26�(a) �.29 .02 �14.53� (a)

� p < .10.� p < .05.


Between 400 and 1000 ms, speakers deployed their gaze to the agent in events with ‘‘easy’’ and‘‘hard’’ agents alike. There was no interaction between Agent codability and Time bin, indicating thatthe slope of fixations in events with ‘‘easy’’ and ‘‘hard’’ agents did not change over time: speakerscontinued fixating ‘‘harder’’ agents more than ‘‘easier’’ agents until 1000 ms.

At the start of the 400–1000 ms time window (i.e., 400–600 ms), speakers were also less likely tofixate agents in ‘‘easier’’ events than ‘‘harder’’ events (a main effect of Event codability). An interactionwith Time bin shows that fixations to the agent subsequently increased more quickly in ‘‘easier’’events than ‘‘harder’’ events.

Fixations between 1000 and 1800 ms (speech onset). Speakers began looking away from the agentapproximately 1000 ms after picture onset, and the cross-over point after which they started fixatingthe patient preferentially occurred approximately 1800 ms into the trial (i.e., around speech onset).

Comparing Fig. 3a and b shows that, at 1000–1200 ms, speakers were less likely to fixate agents in‘‘easy’’ events than in ‘‘hard’’ events (a main effect of Event codability; Table 3c): in addition, speakerswere less likely to fixate ‘‘easy’’ agents in ‘‘easy’’ events but still fixated ‘‘easy’’ agents in ‘‘hard’’ events(producing a weak interaction of Event and Agent codability; Table 3c). There were no interactionswith Time bin, indicating that the decline in agent-directed fixations after 1000 ms was comparableacross event categories. However, since the peak in fixations to the agent occurred earlier in ‘‘easy’’events than ‘‘hard’’ events, the shift of gaze to the patient also occurred earlier in ‘‘easy’’ events than‘‘hard’’ events. On the hypothesis that high Event codability favors faster encoding of relational infor-mation in the event (hierarchical incrementality), this result suggests that speakers began addinginformation about the second character to the developing sentence earlier when the relationshipbetween characters was easier to encode than when it was harder to encode.

2.2.4.2. Effects of Agent codability and Lexical primes. Fig. 4a and b shows the timecourse of formulationfor sentences with ‘‘easier’’ and ‘‘harder’’ agents across the three Prime conditions. In each analysis,fixations across conditions were compared with two contrasts.


Fixations between 0 and 400 ms. Again, speakers directed more fixations to ‘‘easier’’ agents than‘‘harder’’ agents within 200 ms of picture onset (a main effect of Agent codability; Table 4a) and thenbriefly looked back to the patient by 400 ms. An interaction with Time bin was observed only in the

Fig. 4. Timecourse of formulation for active sentences in Experiments 1 and 2 with respect to Agent codability (‘‘easy’’ vs.‘‘hard’’ agents) and Prime condition.

Table 4Results of regressions comparing fixations to the agent in active sentences with ‘‘easy’’ and ‘‘hard’’ agents across Prime conditionsin Experiment 1.



(a) 0–400 msIntercept �1.55 .04 �36.64� �1.88 .03 �58.12�

Time bin 8.48 .25 34.35� 9.61 (s) .24 39.44�


Prime condition: neutral primes vs. other primes .10 .05 1.86� �.18 .04 �4.56�

Prime condition: agent primes vs. patient primes �.13 .06 �2.08� �.42 .05 �8.81�

Time bin � Agent codability – – – �2.09 .30 �6.96�

Time bin � Prime condition: neutral primes vs. other primes – – – 1.77 .29 6.11�

Time bin � Prime condition: agent primes vs. patient primes – – – 2.39 .36 6.72�

(b) 400–1000 msIntercept �.09 .03 �2.79� �.29 .03 �8.81�

Time bin 1.57 .12 13.36� 2.53 (s) .09 27.03�


Prime condition: neutral primes vs. other primes .06 (s) .05 1.19 .05 .02 2.38�

Prime condition: agent primes vs. patient primes .15 (s) .06 2.43� .27 .03 9.99�

Time bin � Agent codability – – – .37 .11 3.35� (a)

(c) 1000–1800 msIntercept .54 .03 16.60� .77 .03 30.00�

Time bin �1.06 .08 �14.05� �1.53 .06 �25.07�


Prime condition: neutral primes vs. other primes �.16 .04 �4.31� �.19 .02 �9.42�

Prime condition: agent primes vs. patient primes �.03 .05 �.74 �.12 .02 �5.10�

� p < .05.� p < .10.


by-item analysis, showing that fixations to ‘‘easier’’ agents rose somewhat more steeply over time inthis time window than fixations to ‘‘harder’’ agents.

Supporting linear incrementality, there were also more fixations to the agent after agent andpatient primes (‘‘other’’ primes) than after neutral primes (the first contrast for Prime condition inthe by-participant analysis) and more fixations to the agent after agent primes than patient primes(the second contrast for Prime condition in the by-participant analysis). The by-item analysis addi-tionally showed that fixations to agents increased more rapidly after ‘‘other’’ primes than after neutralprimes and more rapidly after patient primes than agent primes (the first and second contrast respec-tively in the interaction of Prime condition with Time bin). There was no interaction between Primecondition and Agent codability.

Fixations between 400 and 1000 ms. After fixating ‘‘easier’’ agents preferentially before 400 ms,speakers began looking away from ‘‘easier’’ agents. At 400–600 ms, there were thus fewer fixationsto ‘‘easy’’ agents than ‘‘hard’’ agents (resulting in a main effect of Agent codability; Table 4b), and thisdifference persisted over the entire 400–1000 ms time window (an interaction with Time bin wasobserved only in the by-item analysis).

An effect of Prime condition was present in this time window as well. After fixating agents moreoften after agent primes before 400 ms, speakers were also somewhat more likely to look away fromthe agent after agent primes and patient primes than after neutral primes (the first contrast for Primecondition; Table 4b). This effect was mainly driven by a difference between the agent prime andpatient prime conditions (the second contrast for Prime condition) and shows that priority encodingof a character agent before 400 ms was followed by a larger shift away from this character after400 ms. Overall, this pattern demonstrates a strong tendency for character-by-character encodingduring formulation of the target sentences. There were no interactions of Prime condition with Agentcodability or Time bin.

Fixations between 1000 and 1800 ms (speech onset). ‘‘Easy’’ agents were faster to encode, so speakerswere less likely to fixate ‘‘easy’’ agents than ‘‘hard’’ agents at 1000–1200 ms. There was no interaction


with Time bin, so the difference between ‘‘easy’’ and ‘‘hard’’ agents persisted across the entire timewindow. In addition, there were fewer fixations to agents after agent primes and patient primes(‘‘other’’ primes) than after neutral primes (the first contrast for Prime condition; Table 4c), suggestingthat priming of either character resulted in an earlier shift of gaze to the patient. A differences infixation patterns after agent primes and patient primes was reliable only in the by-item analyses(the second contrast for Prime condition), showing fewer fixations to agents after agent primes. Therewere no interactions with Agent codability or Time bin.

2.3. Discussion

Experiment 1 showed that sentence form was influenced in different ways by non-relational andrelational variables and that the timecourse of formulation reflected these differences.

On the one hand, there was an expected effect of character codability and lexical priming on sen-tence form: speakers produced accessible characters (‘‘easy’’ agents and patients) before less accessi-ble characters in their sentences. This confirms that ease of naming can determine the suitability ofindividual characters for starting points and is broadly consistent with linear incrementality. On theother hand, comparing the agent and patient prime conditions against the neutral prime conditionshows that priming effects after agent and patient primes were asymmetrical: agent primes did notreliably increase production of active sentences whereas patient primes reduced the probability ofselecting an active structure. Thus manipulating the accessibility of a character that speakers normallyproduce in object position (the patient) produced a larger change than manipulating the accessibilityof a character that is more often selected as the sentence subject (the agent). An interaction with Agentcodability additionally showed that the influence of patient primes was stronger in events withharder-to-name agents than easier-to-name agents: speakers were more likely to postpone produc-tion of a ‘‘hard’’ agent than an ‘‘easy’’ agent when the patient had been primed.

Given that stronger priming effects are normally obtained for less frequent than more frequentlexical items and for dispreferred than preferred structures (see Ferreira & Bock, 2006), the resultsdemonstrate that speakers have a strong preference for treating agents as ‘‘default’’ subjects. Theimplication of this result for theories of formulation is that accessibility effects are modulated byhigher-level, relational information: in order to prioritize encoding of the agent, speakers first hadto identify the two characters in terms of their event roles. The degree to which identification of agentsrequires extensive encoding of event gist is debatable because of lower-level perceptual correlates of‘‘agenthood’’ (Hafri et al., 2012); nevertheless, selection of starting points appears to be sensitive to acombination of non-relational and relational variables.

Consistent with this conclusion is also the effect of Event codability on sentence form. The influ-ence of patient primes was stronger in ‘‘harder’’ than ‘‘easier’’ events: the patient primes reducedthe likelihood of selecting the preferred active structure to describe ‘‘hard’’ events, suggesting thatincreasing the accessibility of patient names increased their suitability for starting points in caseswhere properties of the event did not facilitate selection of a starting point on conceptual grounds.Character accessibility played a smaller role in ‘‘easier’’ events, or events where speakers could initiateproduction without relying on properties of the two characters to select a starting point. The directionof this effect is consistent with Kuchinsky and Bock’s (2010) finding that drawing attention to onecharacter is less likely to bias assignment of this character to subject position in ‘‘easier’’ than ‘‘harder’’events.

More importantly, we tested whether these differences in non-relational and relational encodingalso produced different patterns of eye movements across items and conditions before speech onset.The first test of incrementality is the analysis of first fixations, and the results were consistent withlinear incrementality: the first-fixated character was more likely to become the sentence subject thanthe other character. This replicates studies using cuing paradigms to test the effect of gaze shifts onsentence form (Gleitman et al., 2007, Kuchinsky & Bock, 2010) without a direct manipulation of per-ceptual salience and attention capture.

The second test of incrementality is the analysis of eye movements to event characters throughoutthe formulation process. Timecourse analyses showed a combination of accessibility effects andeffects of relational variables on formulation. Eye movements in the first time window (0–400 ms)


showed immediate sensitivity to properties of the agents: speakers directed their gaze towards ‘‘eas-ier’’ agents and away from ‘‘harder’’ agents. Importantly, effects of Agent codability were modulatedby the ease of encoding event gist (Event codability): Agent codability had a weaker influence in‘‘easy’’ events than ‘‘hard’’ events, suggesting that properties of the event can constrain the effectsof character-specific variables. This shows higher sensitivity to relational than non-relational informa-tion, consistent with hierarchical incrementality.

Fast encoding of an ‘‘easy’’ agent before 400 ms was then followed by a shift of attention to thepatient around 400 ms. In other words, speakers did not continue fixating the subject character after400 ms as predicted by the strong version of linear incrementality (Gleitman et al., 2007), but system-atically shifted their gaze back to the patient. This type of character-by-character encoding is consis-tent with a weaker version of linear incrementality, where speakers do attempt to encode informationabout both characters early in the formulation process but, crucially, they encode this informationsequentially. Thus the rise and fall of fixations to the agent after 400 ms was also predicted by Agentcodability: fixations to agents were generally delayed after 400 ms if agents attracted more attentionbefore 400 ms, and vice versa. Specifically, formulation in events with ‘‘harder’’ agents showed thatthere is a benefit to distributing attention more evenly between the two characters before 400 ms: for-mulation after 400 ms continued with rapid, preferential encoding of the agent. Importantly, shifts ofgaze to the agent after 400 ms and away from the agent after 1000 ms were also influenced by Eventcodability: as predicted by hierarchical incrementality, speakers began fixating the patient earlier inhigher-codability events than lower-codability events.

As expected, the lexical primes produced analogous effects to Agent codability: within 400 ms ofpicture onset, speakers directed more fixations to the agent after agent primes than after patientprimes and neutral primes. This shows that the agent primes selectively influenced encoding of theagent character and that they increased the likelihood of speakers prioritizing encoding of this char-acter (non-relational information) shortly after picture onset. A shift of gaze away from the agent wasthen observed after 400 ms, confirming the tendency to encode sentences character by character afterpriming non-relational information. Finally, after 1000 ms speakers looked at the agent for less timeafter agent and patient primes than neutral primes, and thus shifted their gaze to the patient earlierwhen either character had been primed than when the primes mentioned an unrelated character.

Taken together, the results show a direct link between the ease of encoding non-relational and rela-tional information and the timecourse of formulation, both during the early deployment of attentionto the subject character and then the deployment of attention to the object character around speechonset. Effects of event codability generally favor hierarchical incrementality, while pervasive effects ofcharacter accessibility on formulation – originating both in pre-existing properties of the event char-acters and in experimentally manipulated accessibility – generally favor linear incrementality. Thenext experiment tested whether experimentally facilitating relational encoding via structural primingproduces a shift in planning in the opposite direction to that obtained in Experiment 1 with a manip-ulation of the ease of non-relational encoding.

3. Experiment 2

Speakers completed a similar task in the second experiment. Target pictures included a nearlyidentical set of two-character transitive events as that of Experiment 1. Formulation of activesentences was again compared from picture onset until speech onset with respect to one variableinfluencing encoding of individual characters (character codability) and one variable influencing rela-tional encoding (event codability). Effects of character codability and event codability were expectedto replicate Experiment 1.

On the hypothesis that relational encoding also depends on the ease of generating a syntacticstructure, the ease of structural assembly was manipulated by exposing speakers to three types ofstructural primes before target trials. On one third of all prime trials, speakers saw a picture of a tran-sitive event that was accompanied by a recorded active sentence, and on one third of all trials, theysaw the same picture accompanied by a recorded passive description. Active and passive syntaxwas thus either primed or unprimed. On the remaining third of prime trials, speakers saw pictures


where multiple referents were engaged in a joint action and heard an intransitive sentence (e.g., Thecouple are roller-skating). This condition served as a baseline to assess the overall likelihood of usingactive and passive syntax.

If the ease of structural assembly influences the timecourse of sentence formulation, speakersshould be more likely to engage in hierarchically incremental planning when using primed structuresthan unprimed structures. Specifically, if fast generation of an ‘‘easy’’ structure facilitates encoding ofrelational information about the event (i.e., the relationship between two characters), fixations toagents and patients should diverge more slowly in the first 400 ms of picture inspection in primed sen-tences compared to unprimed sentences. This is analogous to the effect of Event codability on earlyformulation. Having encoded relational information in primed sentences before 400 ms, fixations tothe two characters after 400 ms should then show evidence of top-down structural guidance: speakersshould direct their gaze to the agent more quickly in primed than unprimed sentences after 400 msand should begin shifting their gaze to the patient earlier in primed than primed sentences aroundspeech onset.

3.1. Method

3.1.1. ParticipantsEighty-four native speakers of Dutch (mostly university students; 64 female) from the Nijmegen

area participated for payment. Seven participants were replaced because of low response rates andseven because of technical problems. The sample was larger than in Experiment 1 because partici-pants also took part in a second, unrelated study.

3.1.2. Materials, design, and procedureAs in Experiment 1, there were four types of trials: target trials, prime trials, filler trials, and word

trials. On target trials, participants saw pictures of two-character transitive events (26 pictures usedin Experiment 1 and 4 new pictures) and were asked to describe them in one sentence. There were21 items with animate agents (13 items with human agents, 8 with animal agents), and 9 with inan-imate agents. Twenty-two items had animate patients (19 items had human patients, 3 had animalpatients) and 8 had inanimate patients.

Target pictures were preceded by three types of prime trials. In the active and passive prime con-ditions, speakers saw new pictures of two-character transitive events accompanied by a recordedactive or passive description. In the neutral prime condition, they saw pictures of two-character (ormulti-character) intransitive events accompanied by a recorded intransitive description.

The design included one three-level factor (Prime condition: active primes, passive primes, neutralprimes). Two versions of each target picture were created to counterbalance the location of agents andpatients in each picture on the left and right hand-side of the screen, but all analyses collapsed acrossthis factor. The procedure and list structure were analogous to Experiment 1.

3.1.3. Sentence scoringThe same scoring criteria were applied as in Experiment 1. Two items were excluded from the anal-

yses because they elicited a very low number of scorable responses. Responses were also excluded ifthe first fixation in the trial did not fall on the fixation point at the top of the screen (144 trials), iflatencies were longer than 5.5 s and longer than 3 standard deviations away from the grand mean(28 sentences; the 5.5 s cutoff is higher than in Experiment 1 because sentence onsets were on averagelonger than in the first experiment). After applying these criteria, there were 1405 trials (.68 actives,.32 full passives) left for the analyses of structure choice and for the timecourse analyses. Excludingdisfluent responses left 1334 trials for the analysis of speech onsets.

3.1.4. Character codability and Event codability scoringCodability ratings were calculated as in Experiment 1. Again, Event codability was not correlated

with either Agent or Patient codability (r = .18 and �.27, ns., respectively), confirming that encodingof the relational structure of an event did not depend on fast identification and naming of individualcharacters. Event codability ratings did not differ across Prime conditions (all ps > .9), showing that the


structural primes did not influence speakers’ verb choice and thus did not contribute further to vari-ability in event descriptions. Agent and Patient codability were again positively correlated (r = .45,p < .05). Since Patient codability (residualized on Agent codability) did not reliably influence structurechoice, this factor was excluded from all analyses.

3.1.5. AnalysesAnalyses were conducted as in Experiment 1 and considered effects of First fixations (Section 3.2.1)

and structural primes on sentence form (Section 3.2.2) across items and conditions, differences inspeech onsets across items and conditions (Section 3.2.3), and the timecourse of formulation(Section 3.2.4) for active sentences.

3.2. Results

3.2.1. First fixations and sentence formThe majority of first fixations were directed to the agent (.68), as in Experiment 1, and the distri-

bution of first fixations was influenced by structural primes: speakers directed fewer fixations to theagent at picture onset after active primes (.64) than after other primes (neutral and patient primes; .70and .71 respectively, b = �.50, z = �3.03). The neutral and passive prime conditions did not differ(b = .05, z = .25). Thus unlike the lexical primes in Experiment 1, the influence of structural primeson visual inspection of a pictured event was not to direct speakers’ gaze to the agent after activeprimes and to the patient after passive primes: in other words, structural primes did not primeselection of a particular character as a starting point.

First fixations were also weaker predictors of sentence form than in Experiment 1 (Fig. 1b and c).Fig. 1b shows that the degree to which first fixations influenced structure choice was modulated bythe structural primes. Speakers produced more active sentences if they looked first at the agent ratherthan at the patient after neutral primes and passive primes; this pattern was reversed after activeprimes, where speakers produced actives after both agent-directed and patient-directed first fixations.In addition, the effect of active primes on structure selection was stronger in ‘‘easier’’ events (Fig. 1c),where speakers produced actives even after first looking at the patient, than in ‘‘harder’’ events, wherespeakers were generally more likely to assign a first-fixated character to subject position. This resultedin a three-way interaction between First Fixations, Prime condition, and Event codability (b = �1.09,z = �2.19, with random by-participant slopes for First Fixations and Prime condition, and randomby-item slopes for Prime condition; the interaction was reliable but did not improve model fit). Inother words, the two variables influencing the ease of relational encoding (Event codability and struc-tural priming) reduced the impact of first fixations on selection of a sentence structure.

3.2.2. Structural primes, codability, and sentence formFig. 2a shows the proportions of active sentences produced in the three Prime conditions. Speakers

produced fewer active sentences after passive primes than after other primes (active and neutralprimes; the first contrast for Prime condition in Table 2). Production of actives after active primesand neutral primes was comparable: relative to the neutral baseline condition, active primes didnot increase likelihood of speakers producing active sentences (the second contrast for Prime condi-tion in Table 2). In other words, the effectiveness of active primes was considerably weaker than theeffectiveness of passive primes, confirming the observation that the magnitude of priming is larger forthe less frequent construction (Ferreira & Bock, 2006).

Agent codability had the expected effect on sentence form: speakers produced more active sen-tences beginning with ‘‘easy’’ agents than ‘‘hard’’ agents (.71 vs. .61). Importantly, Agent codabilityinteracted with Prime condition (Fig. 2b; Table 2). The first contrast for this interaction shows no dif-ference between production of actives in the passive condition and in the two other conditions initems with ‘‘easy’’ and ‘‘hard’’ agents. However, the second contrast shows a difference between theactive prime condition and neutral prime condition: this is due to the fact that active primes increasedthe likelihood of speakers placing a ‘‘harder’’ agent in subject position. In other words, the effect ofagent accessibility on sentence form was attenuated by structural priming: active primes selectivelyincreased production of active descriptions in items where properties of the agent disfavored selection


of active syntax. The direction of this effect is again consistent with the observation that primingeffects are larger when structures are difficult to produce (‘‘difficulty’’ in this case is defined by theconflict between the preference to begin sentences with agents and the preference to produce lessaccessible referents later).

Structure choice was not sensitive to Event codability (Fig. 2c). Speakers tended to produce moreactive sentences to describe ‘‘harder’’ events, and, while passive primes reduced this tendency,interactions with Prime condition did not reach significance.

3.2.3. Speech onsetsActive sentences were initiated earlier than passive sentences (2029 ms vs. 2131 ms). As in Exper-

iment 1, onsets were sensitive to Agent codability: sentences with ‘‘easier’’ agents were initiated morequickly than sentences with ‘‘harder’’ agents (b = .16, z = 3.51, for the main effect of Agent codability),but this effect was smaller in passive sentences, where agents were produced in object position(b = .08, z = 2.18, for the interaction of Sentence type with Agent codability). Thus speakers likelyattempted to encode agents as sentence subjects by default, but demonstrated more sensitivity toproperties of the second character than in Experiment 1.

Speech onsets differed across Prime conditions only in active sentences. Onsets were longer afterpassive primes than after active and neutral primes combined (b = .08, z = 2.98); onsets after activeand neutral primes did not differ (b = .01, z = .22). Onsets in passive sentences did not vary by condi-tion, but interactions of Sentence type (active vs. passive) with Prime condition did not reachsignificance.

3.2.4. Timecourse analysesAs in Experiment 1, speakers began formulation of active sentences by fixating agents preferen-

tially within 200 ms of picture onset and then briefly directing their gaze to the patient by 400 ms.They redeployed attention to agents after 400 ms and continued fixating this character until speechonset.

Analyses were carried out in three steps. The first analysis compared formulation of sentences forevents varying in Event codability and Agent codability (Section 3.2.4.1). The second analysis exam-ined formulation of sentences with ‘‘easy’’ and ‘‘hard’’ agents across Prime conditions (Section 3.2.4.2),and the third analysis examined formulation of sentences describing ‘‘easy’’ and ‘‘hard’’ across Primeconditions (Section 3.2.4.3). Three time windows were chosen for examination within each analysis:0–400 ms, 400–1000 ms (showing an increase in agent-directed fixations), 1000–2200 ms (i.e., speechonset; showing a decrease in agent-directed fixations).

3.2.4.1. Effects of Agent codability and Event codability. Fixations between 0 and 400 ms. Fig. 3c and dshows the timecourse of formulation for descriptions of ‘‘easy’’ and ‘‘hard’’ events with ‘‘easy’’ and‘‘hard’’ agents. The best-fitting model included a three-way interaction between Event codability,Agent codability, and Time bin (Table 5a). As in Experiment 1, speakers generally preferred to fixate‘‘easy’’ agents at and shifted their gaze away from ‘‘hard’’ agents in search of an alternative startingpoint (producing an interaction of Agent codability with Time bin), consistent with linear incremen-tality. Event codability had the opposite effect: speakers distributed their gaze more evenly betweenagents and patients in ‘‘easy’’ events but directed more fixations to agents than patients in ‘‘hard’’events. Critically, the three-way interaction shows that the effect of Agent codability depended onproperties of the event. The difference between fixations directed to ‘‘easy’’ and ‘‘hard’’ agents wasrelatively small in ‘‘easy’’ events (Fig. 3c) and larger in ‘‘hard’’ events (Fig. 3d): here, fixations to aneasy-to-name agent rose more quickly than to a harder-to-name agent. Thus speakers showed moresensitivity to properties of the agent when the relational structure of the event was harder to encode,which is broadly consistent with hierarchical incrementality.

Interestingly, as in Experiment 1, the shift of gaze away from the agent before 400 ms in items with‘‘easy’’ agents suggests that fast selection of a starting point was likely insufficient to continue formu-lation without encoding information about the patient.

Fixations between 400 and 1000 ms. Following from differences in the distribution of fixations acrossitems observed immediately after picture onset, speakers were less likely to fixate ‘‘easy’’ agents than

Table 5Results of regressions comparing fixations to the agent in active sentences describing ‘‘easier’’ and ‘‘harder’’ events with ‘‘easier’’and ‘‘harder’’ agents in Experiment 2.



(a) 0–400 msIntercept �1.84 .03 �61.61� �2.03 .01 �190.52�

Time bin 10.52 (s) .18 59.40� 11.82 (s) .11 109.22�



Time bin � Event codability 2.67 .23 11.42� .94 .13 7.01� (a)Time bin � Agent codability �1.76 .24 �7.32� �1.27 .11 �11.18� (a)Event codability � Agent codability �.78 .08 �9.85� �.18 .01 �12.59� (a)Time bin � Event cod. � Agent cod. �2.67 .49 �5.46� �.52 .15 �3.53� (a)

(b) 400–1000 msIntercept �.15 .02 �9.34� �.29 .02 �15.52�

Time bin 1.75 (s) .07 26.48� 2.80 .03 80.52�

Event codability .24 .03 8.70� �.01 .02 �.29Agent codability .23 .03 7.97� .11 .02 5.55�

Time bin � Event codability �1.70 .11 �14.87� �.63 .04 �14.81� (a)Time bin � Agent codability 1.22 .12 10.06� 1.12 .04 28.49�

Event cod. � Agent codability .58 .04 13.77� �.06 .02 �2.32� (a)

(c) 1000–2200 msIntercept .62 .02 30.82� .94 .01 72.58�

Time bin �.91 (s) .03 �29.32� �1.19 (s) .01 �88.23�

Event codability �.07 .03 �2.35� �.08 .02 �4.75�


Time bin � Event codability .35 .05 6.37� .29 .02 17.68� (a)Event codability � Agent codability �.30 .04 �8.08� �.06 .01 �4.40� (a)

� p < .05.


‘‘hard’’ agents and less likely to fixate agents in ‘‘easy’’ than ‘‘hard’’ events at 400–600 ms (main effectsof Agent and Event codability respectively; Table 5b). The two factors interacted: the difference infixations directed to ‘‘easy’’ and ‘‘hard’’ agents was again larger in ‘‘hard’’ events than in ‘‘easy’’ events.As there was no three-way interaction with Time bin, this difference persisted across the entire timewindow.

Individual interactions of Agent and Event codability with Time bin showed that the rise infixations to the agent after 400 ms depended both on properties of the agents and of the events.The interaction of Agent codability with Time bin shows that the difference in fixations to the ‘‘easy’’and ‘‘hard’’ agents increased over time. As in Experiment 1, the interaction of Event codability withTime bin shows that speakers directed their gaze to the subject character more rapidly in ‘‘easy’’events than ‘‘hard’’ events.

Fixations between 1000 and 2200 ms. Speakers began looking away from the agent approximately1000 ms after picture onset and switched their gaze to the patient approximately 1 s later (aroundspeech onset). At 1000–1200 ms, speakers were generally less likely to fixate ‘‘easy’’ agents than‘‘hard’’ agents, and more likely to fixate agents in ‘‘easy’’ events than ‘‘hard’’ events; the two factorsinteracted in the by-participant analysis (Table 5c).

There was no three-way interaction with Time bin, indicating that this difference persisted acrossthe entire time window. As a result, speakers also shifted their gaze to the patient most quickly in‘‘easy’’ events with ‘‘easy’’ agents. Importantly, the strongest predictor of the timing of the gaze shiftfrom agents to patients was Event codability: speakers looked to the patient earlier in ‘‘easier’’ eventsthan ‘‘harder’’ events (an interaction of Event codability with Time bin). Consistent with hierarchicalincrementality, this result suggests that speakers were able to begin adding the second character tothe sentence earlier in items where the event gist was easier to encode.


3.2.4.2. Effects of Agent codability and Structural priming. Fixations between 0 and 400 ms. Fig. 4c and dshows the timecourse of formulation for items with ‘‘easier’’ and ‘‘harder’’ agents across Prime condi-tions. Again, speakers were more likely to fixate ‘‘easy’’ agents than ‘‘harder’’ agents across the entiretime window: the effect of Agent codability was reliable in the 0–200 ms time bin (main effect ofAgent codability in Table 6a) and was somewhat larger in the 200–400 ms time bin (an interactionof Agent codability with Time was observed in the by-item analysis).

As predicted by hierarchical incrementality, early fixations to the agent were influenced by struc-tural priming. Speakers directed fewer fixations to the agent after active primes than after otherprimes at 0–200 ms (neutral and passive primes; the first contrast for Prime condition in Table 6a);there was no interaction with Time bin, indicating that this difference persisted into the 200–400 ms time bin. The distribution of agent-directed fixations did not differ after neutral and passiveprimes (the second contrast for Prime condition). The priming effect also did not vary systematicallywith properties of the agent (no interaction with Agent codability).

Fixations between 400 and 1000 ms. Having shifted their attention away from ‘‘easy’’ agents by400 ms, speakers were less likely to fixate ‘‘easy’’ agents than ‘‘hard’’ agents at 400–600 ms (a maineffect of Agent codability; Table 6b). Fixations to ‘‘easy’’ agents subsequently also rose less steeplythan fixations to ‘‘hard’’ agents (an interaction of Agent codability with Time bin).

Importantly, the probability of fixating the agent was higher after active primes and passive primesthan neutral primes at 400–600 ms (the first contrast for Prime condition), and this difference

Table 6Results of regressions comparing fixations to the agent in active sentences with ‘‘easy’’ and ‘‘hard’’ agents across Prime conditionsin Experiment 2.



(a) 0–400 msIntercept �1.78 .04 �48.58� �2.01 .02 �94.38�

Time bin 10.08 (s) .22 45.59� 11.11 (s) .17 66.06�


Prime condition: active primes vs. other primes .20 .04 5.54� .30 (s) .02 18.39�

Prime condition: neutral primes vs. passive primes .05 .04 1.20 .01 (s) .02 .33Time bin � Agent codability – – – �2.09 .16 �12.91�

(b) 400–1000 msIntercept �.13 .02 �6.84� �.32 .03 �10.09�

Time bin 1.66 (s) .08 20.61� 2.81 (s) .07 39.00�


Prime condition: neutral primes vs. other primes .02 .04 .52 .07 .02 3.59�

Prime condition: active primes vs. passive primes �.21 .04 �5.07� �.26 .02 �10.99�

Time bin � Agent codability 1.03 .14 7.28� .88 .07 11.73�

Time bin � Prime condition: neutral primes vs. other primes .75 .14 5.39� .80 .09 9.32�

Time bin � Prime condition: active primes vs. passive primes .93 .16 5.63� 1.07 .10 10.52�

(c) 1000–2200 msIntercept .56 .02 23.80� .83 .02 41.69�

Time bin �.81 .04 �22.10� �1.08 (s) .02 �45.93�


Prime condition: active primes vs. other primes .05 .02 2.71� �.07 .02 �3.83�

Prime condition: neutral primes vs. passive primes �.02 .02 �.69 .10 .02 4.69�

Agent codability � Prime condition: active primes vs. otherprimes

�.06 .04 �1.54 �.11 .01 �10.95�

Agent codability � Prime condition: neutral primes vs. passiveprimes

.27 .05 5.48� .14 .01 11.83�

Time bin � Prime condition: active primes vs. other primes – – – .37 .03 11.64�

Time bin � Prime condition: neutral primes vs. passive primes – – – �.33 .04 �8.74�

� p < .10.� p < .05.


increased over time (the first contrast in the interaction of Prime condition with Time bin), suggestingpossible facilitation from exposure to a transitive sentence or a transitive-event conceptual structure.In addition, there were also more fixations to the agent after active primes than passive primes at 400–600 ms (the second contrast for Prime condition), although fixations to the agent then rose moresharply after passive primes (the second contrast in the interaction of Prime condition with Timebin). The overall pattern is thus different from Experiment 1, where fixations to the agent decreasedafter agent primes relative to other primes, and shows evidence of guidance from a larger frameworkduring linguistic encoding.

Fixations between 1000 and 2200 ms (speech onset). At 1000–1200 ms, speakers were less morelikely to fixate ‘‘easy’’ agents than ‘‘hard’’ agents (a main effect of Agent codability; Table 6c). The ratesat which fixations to the agent decreased over time in items with ‘‘easy’’ and ‘‘hard’’ agents did notdiffer (no interaction of Agent codability with Time bin).

Differences across Prime conditions were observed in this time window as well. The by-participantanalysis shows that there were fewer fixations to the agent after active primes than other primes at1000–1200 ms (the first contrast for Prime condition), and the absence of an interaction with Timebin suggests that this difference persisted across the entire time window. By comparison, the by-itemanalysis shows a steeper decline in agent-directed fixations after active primes than after other primes(the first contrast in the interaction of Prime condition with Time bin). Together, the two analyses sug-gest that speakers spent less time fixating agents in structurally primed (active-primed) sentences. Adifference between passive primes and neutral primes was observed only in the by-item analysis.

In addition, priming effects were sensitive to properties of the agents. The first contrast in the inter-action of Agent codability with Prime condition shows that, at 1000–1200 ms, there were somewhatmore fixations to agents after active primes than other primes in items with ‘‘hard’’ agents (the effectreached significance in the by-item analysis). The second contrast in the interaction of Agent codabil-ity with Prime condition shows that, at 1000–1200 ms, there were more fixations to agents afterpassive primes than neutral primes in items with ‘‘hard’’ agents.

3.2.4.3. Effects of Event codability and Structural priming. Fixations between 0 and 400 ms. Fig. 5a and bshows the timecourse of formulation for sentences describing ‘‘easy’’ and ‘‘hard’’ events across Primeconditions. There were more fixations to agents in ‘‘hard’’ events than ‘‘easy’’ events at 0–200 ms; in‘‘easy’’ events, speakers were less likely to fixate either character preferentially (a main effect of Eventcodability; Table 7a). Fixations to the agent then increased more quickly in ‘‘hard’’ events than in‘‘easy’’ events by 400 ms (producing an interaction of Event codability with Time bin).

Again, speakers directed fewer fixations to the agent after active primes than after neutral and pas-sive primes in the 0–200 ms time bin (the first contrast for Prime condition), and this effect persistedinto the 200–400 ms time bin (there was no interaction with Time bin). In addition, the reduction inagent-directed fixations with structural priming was larger in ‘‘easier’’ than ‘‘harder’’ events (the firstcontrast in the interaction between Prime condition and Event codability).

Speakers were also somewhat more likely to fixate agents after passive primes than neutral primes(the second contrast for Prime condition), particularly in ‘‘harder’’ events (the second contrast in theinteraction between Prime condition and Event codability).

Fixations between 400 and 1000 ms. At 400–600 ms, speakers were less likely to fixate agents in‘‘easy’’ events than ‘‘hard’’ events (a main effect of Event codability in the by-participant analysis;Table 7b), but fixations to the agent then rose more quickly in ‘‘easy’’ events than ‘‘hard’’ events(resulting in an interaction between Event codability and Time bin). There were also more agent-directed fixations after active primes and passive primes than after neutral primes at 400–600 ms(the first contrast for Prime condition), and fixations to the agent rose more quickly in these conditionsover time (the first contrast in the interaction of Prime condition with Time bin). Additionally, speak-ers were more likely to fixate the agent after active primes than passive primes at 400–600 ms, butfixations to the agent then increased more quickly after passive primes than after active primes(the second contrast in the interaction of Prime condition with Time bin).

Fixations between 1000 and 2200 ms (speech onset). At 1000–1200 ms, speakers were somewhat lesslikely to fixate the agent in ‘‘easy’’ events than ‘‘hard’’ events (a main effect of Event codability in theby-participant analysis; Table 7c). There was no interaction with Time in the by-participant analysis,

Fig. 5. Timecourse of formulation for active sentences in Experiment 2 with respect to Event codability (‘‘easy’’ vs. ‘‘hard’’events) and Prime condition.


suggesting that this difference persisted across the entire time window and resulted in an earlier shiftof gaze to the patient in ‘‘easy’’ events than ‘‘hard’’ events. This interaction was reliable in the by-itemanalysis, indicating that the decline in agent-directed fixations after 1000 ms was faster in ‘‘easy’’events than in ‘‘hard’’ events. Together, the two analyses show that speakers fixated agents for lesstime when the gist of the event was easy to encode than when it was harder to encode.

Finally, speakers were less likely to look at the agent after active primes than after other primes at1000–1200 ms (the first contrast for Prime condition) and then shifted their gaze to the patient earlierafter active primes (the first contrast in the interaction of Prime condition with Time bin). They werealso somewhat more likely to shift their gaze to the patient earlier after neutral primes than passiveprimes (the second contrast in the interaction of Prime condition with Time bin).

3.3. Discussion

Experiment 2 showed effects of non-relational and relational variables on both sentence form andsentence formulation that were similar to those used in Experiment 1. With respect to sentence form,the results showed the expected robust effects of character accessibility and structural priming. Prop-erties of agents were again strong predictors of sentence form, consistent with linear incrementality.The structural priming manipulation showed that sentence form was also influenced by changes in theease of deploying abstract structure-building procedures, and again, the primes differed in their prim-ing ability: speakers produced a comparable number of active sentences after active primes and neu-tral primes, whereas only passive primes reliably reduced production of actives. Effects of the activeprimes were limited to items with ‘‘harder’’ agents, or items where properties of the agent favoredselection of a passive structure rather than the preferred active structure. Thus as in Experiment 1,

Table 7Results of regressions comparing fixations to the agent in active sentences describing ‘‘easier’’ and ‘‘harder’’ events across Primeconditions in Experiment 2.



(a) 0–400 msIntercept �1.76 .04 �48.61� �2.01 .02 �102.14�

Time bin 9.77 (s) .21 45.88� 11.11 (s) .18 63.44�


Prime condition: active primes vs. other primes .19 .03 5.73� .28 .02 16.85�

Prime condition: neutral primes vs. passive primes .05 .04 1.38 .04 .02 1.94�

Event codability � Prime condition: active primes vs. otherprimes

�.46 .07 �6.93� �.21 .02 �10.76� (a)

Event codability � Prime condition: neutral primes vs.passive primes

.26 .08 3.36� .10 .02 4.38� (a)

(b) 400–1000 msIntercept �.10 .02 �5.72� �.32 .03 �9.68�

Time bin 1.33 .07 19.41� 2.76 (s) .08 34.88�

Event codability .11 .03 3.25� .01 .04 .23Prime condition: neutral primes vs. other primes .10 .04 2.93� .07 .02 3.43�

Prime condition: active primes vs. passive primes �.20 .04 �4.92� �.26 .02 �11.11�

Time bin � Event codability �.74 .14 �5.41� �.32 .10 �3.32� (a)Time bin � Prime condition: neutral primes vs. other primes .46 .14 3.15� .86 .09 9.87�

Time bin � Prime condition: active primes vs. passive primes .73 .17 4.37� 1.07 .10 10.37�

(c) 1000–2200 msIntercept .55 .03 21.98� .85 .02 39.77�

Time bin �.80 (s) .04 �20.53� �1.08 (s) .02 �47.41�

Event codability .17 .02 8.76� �.03 .03 �1.25Prime condition: active primes vs. other primes �.08 .04 �2.17� �.05 .02 �3.01�

Prime condition: neutral primes vs. passive primes .01 .04 .32 .07 .02 3.27�

Time bin � Event codability – – – .31 .03 10.97�

Time bin � Prime condition: active primes vs. other primes .30 .06 4.70� (a) .38 .03 11.62�

Time bin � Prime condition: neutral primes vs. passiveprimes

�.14 .08 �1.79� �.31 .04 �8.31�

� p < .05.� p < .10.


the asymmetry in priming effects is consistent with earlier observations that generation of a frequentstructure is influenced by priming to a lesser degree than generation of an infrequent structure.

More importantly, the timecourse of formulation again showed sensitivity to the ease of encodingnon-relational and relational information. First, the analysis of first fixations showed that the degreeto which speakers began sentences with the first-fixated character depended on higher-level factors.The suitability of a character for subjecthood depended on the ease of encoding the event and the easeof constructing a suitable sentence structure: first-fixated characters were less likely to become sub-jects in ‘‘easier’’ events than ‘‘harder’’ events and in structurally primed sentences than unprimed sen-tences. Thus overall, the influence of visual information on selection of a starting point was relativelyweak: although speakers may begin sentences with the first-fixated character in subject position(Experiment 1, linear incrementality), sentence form is also the product of more complex interactionsbetween lower-level perceptual factors and higher-level relational factors (Experiment 2, hierarchicalincrementality).

Second, timecourse analyses showed a strong influence of variables influencing encoding of rela-tional information and a weaker effect of variables influencing encoding of non-relational information.Effects of Event and Agent codability (Table 5) were comparable to those in Experiment 1 (Table 3), asthe two experiments used similar items. During early formulation (0–400 ms), Event codability pro-duced effects consistent with hierarchically incremental planning, and, importantly, constrained theeffect of Agent codability. When describing higher-codability events, speakers showed only a small


preference for the agent over the patient, and properties of the agent were weak predictors of the mag-nitude of this preference. In lower-codability events, on the other hand, the pattern of early fixationswas primarily determined by Agent codability: speakers shifted their attention very rapidly to ‘‘easy’’agents and away from ‘‘hard’’ agents. As in Experiment 1, this result suggests that speakers attemptedto select a starting point based on character accessibility when they could not easily select a startingpoint based on their construal of the gist of the event. It also extends Kuchinsky and Bock’s (2010)observations about the influence of relational factors on selection of starting points to the timecourseof sentence formulation.

The benefits of early encoding of event gist carried over to later time windows as well. In higher-codability events, speakers directed their attention to the agent relatively quickly after 400 ms. Bycomparison, the strong preference to fixate the agent in lower-codability events before 400 msresulted in a less consistent pattern of fixations: rapid shifts of attention to the agent within400 ms of picture onset were followed by an extended time window where speakers fixated thepatient (as in Experiment 1, large shifts of attention from one character to another suggest that thetwo characters were encoded sequentially). As a result, agent-directed fixations after 400 ms alsoshowed a joint influence of Event and Agent codability: speakers were able to deploy their attentionto the agent and finally shift their gaze to the patient earlier in ‘‘easier’’ events than in ‘‘harder’’ events(this effect was stronger than in Experiment 1, which showed a main effect of Event codability but nointeraction of Event codability with Time bin).

Critically, the effect of structural primes on formulation was different from the effect of lexicalprimes in Experiment 1: the structural primes produced shifts in planning patterns that resembledthe effect of Event codability on formulation and thus were consistent with hierarchical incremental-ity. As predicted, active primes reduced the proportion of agent-directed fixations within 400 ms ofpicture onset in active sentences, suggesting a very early effect of structural processes on visualinspection of an event. The interaction with Event codability in this time window indicates strongerfacilitation of early relational encoding when both conceptual and linguistic structures were easy togenerate. After active primes, speakers also quickly directed their gaze to the agent after 400 msand to the patient before speech onset. Thus differences across prime conditions show a differencein the way speakers initiated the formulation process (0–400 ms) and then proceeded to encode thetwo characters linguistically (after 400 ms) with and without structural support.

4. General discussion

4.1. Summary

In two experiments, speakers described ‘‘easy’’ and ‘‘hard’’ events with ‘‘easy’’ and ‘‘hard’’ charac-ters after receiving lexical primes (Experiment 1) or structural primes (Experiment 2). Variablesknown to influence sentence form produced the expected effects in both experiments. On the onehand, strong effects of character codability, as well as experimentally manipulated character nameaccessibility in Experiment 1, confirm that speakers prefer to encode accessible characters first andthus build structures that accomodate placement of these characters in subject position (e.g.,Altmann & Kemper, 2006; Bock, 1986b; Ferreira, 1994; Gleitman et al., 2007; Prat-Sala & Branigan,2000). On the other hand, strong effects of event codability in both experiments, as well as experimen-tally manipulated ease of structural assembly in Experiment 2, show that conceptual processes andabstract structural processes attenuate effects of character codability on sentence form. The two setsof results, obtained with similar sets of items, show the influence of two different processes on thegeneration of a sentence structure: one illustrates lexical guidance and the other illustrates theinfluence of relational processes.

These effects originated in different types of incremental planning. Analyses of eye-movementsacross a range of time windows consistently revealed a direct link between the ease of executingnon-relational and relational processes and the way that speakers prepared and assembled differentsentence increments. First, first-fixated characters tended to become sentence subjects but the easeof gist encoding and structural assembly reduced the impact of first fixations on sentence form:


first-fixated characters were less likely to become subjects with structural support. Second, the distri-bution of fixations to the two characters within 400 ms of picture onset also showed opposite effectsof non-relational and relational variables. The ease of encoding individual characters predicted thelikelihood of speakers preferentially fixating one character over the other character, suggesting fastencoding of non-relational information at the outset of formulation. In contrast, the ease of encodingthe conceptual structure of an event and assemblying an abstract syntactic structure determined theextent to which speakers distributed their gaze between two characters more equally, suggestingimmediate sensitivity to relational information as well.

Differences in formulation across items and conditions were also observed between 400 ms and thepoint of gaze shifts to the second character. Overall, fast deployment of gaze to a subject character atthe outset of formulation (0–400 ms) was followed by a shift of gaze away from this character at400 ms (indicating sequential, character-by-character encoding), while distributing attention betweentwo characters at the outset of formulation (0–400 ms) was followed by faster gaze shifts to the sub-ject character after 400 ms. The former was observed in events with ‘‘easy’’ agents and in events withlexically primed agents; the latter was observed in ‘‘easy’’ events and in events that were structurallyprimed. At speech onset, gaze shifts from the agent to the patient followed from the distribution offixations seen in earlier windows and were thus also predicted by properties of the events, propertiesof the agents, and by the lexical and structural primes.

In all comparisons, the two variables that were not manipulated experimentally (event and char-acter codability) and the two variables that were experimentally controlled (ease of lexical and struc-tural encoding in Experiments 1 and 2) produced similar results. Similarity of these effects does notequate the precise mechanisms underlying conceptual and linguistic encoding, but it confirms thatprocessing differences relevant for formulation are between the class of processes that influenceencoding of discrete, non-relational pieces of information (individual characters) and the class of pro-cesses that influence encoding of relationships between characters. Thus in the transition fromthought to speech, variability in formulation can be traced back to the encoding of two qualitativelydifferent types of information, and specifically, to the speed with which these encoding operations canbe completed (also see Konopka, 2012).

4.2. Flexibility in incremental message and sentence formulation

The combined effects of non-relational and relational variables as well as speakers’ sensitivity tothe ease of carrying out these processes suggests that, while these variables systematically influenceformulation, production may be neither strictly linearly incremental nor strictly hierarchically incre-mental. Indeed, the findings of Experiment 1 and 2 are more consistent with weaker versions of bothlinear and hierarchical incrementality rather than with a deterministic, inflexible planning process.For example, with respect to selection of sentence structure, speakes may select first-fixated charac-ters as starting points, but preferential encoding of agents over patients suggests that the assignmentof characters to the subject slot also depends on relational biases. Similarly, accessible characters aremore likely to become subjects than less accessible characters, but these effects also depend on theinfluence of relational variables. With respect to the timecourse of formulation, non-relational andrelational variables jointly influenced the early distribution of fixations to event characters and thetiming of gaze shifts from one character to another. For example, early shifts of gaze to accessibleagents in active sentences (0–200 ms) showed an early effect of non-relational variables, but rapidshifts of gaze to patients by 400 ms showed that speakers do not necessarily continue encoding thatcharacter preferentially before speech onset. Evidence for one form of planning being preferred overthe other was only obtained in two timecourse analyses: event codability largely determined the dis-tribution of early agent-directed fixations in Experiments 1 and 2, with speakers falling back on linearincrementality only when relational encoding was harder to complete.

Overall, the results support the proposal of a continuum of incremental planning that permits shiftsin planning strategies from sentence to sentence. The two experimental manipulations highlightedthese shifts directly: lexical priming in Experiment 1 produced a shift to the linear end point of thecontinuum (Gleitman et al., 2007), while structural priming in Experiment 2 produced a shift to thehierarchical end point of this continuum (Bock et al., 2004; Kuchinsky & Bock, 2010; Kuchinsky


et al., 2011). In sum, the production system allows for the order of encoding operations to be flexible:production may begin both with encoding of single characters and with the formulation of a ‘‘thought’’or idea – something akin to a proposition – in different contexts.

Comparing across experiments, the principle behind this flexibility appears to be a general prefer-ence for completing easier processes before harder processes. This resembles Levelt’s (1989) minimalload principle at the discourse level (also see Ferreira & Henderson, 1998): completing easy processesbefore hard processes lightens the load on the production system and enables speakers to quicklybegin and complete encoding of individual increments. For example, reliance on activation patternsof individual words to select a starting point can be beneficial in so far as it allows speakers to produceaccessible words first and quickly shift processing resources to the next increment (Ferreira & Swets,2002; also see Ferreira, 1996). In contrast, prioritizing encoding of relational information can be ben-eficial in so far as a larger message framework can provide top-down guidance to rapidly bind individ-ual increments of a sentence (e.g., individual words) into a full utterance. Given that the processingdemands of production in every-day situations can change from context to context (as they did inthese experiments), minimizing processing load may be a ubiquitous planning strategy. Earlier worksuggested that flexibility can benefit speed as well as fluency of speech (e.g., Ferreira & Swets, 2002;Levelt, 1989; Wagner et al., 2010), so the specific balance between linearly and hierarchically incre-mental planning may reflect rapid (and implicit) weighing of the different advantages conferred bythese planning strategies in each production context separately.

4.3. Relationship between message-level and sentence-level structures

Sensitivity to differences in ease of encoding during formulation bears on two questions relevantfor most production models: questions about the representation of conceptual and linguistic struc-tures and thus questions about information flow in the production system. Similarity in the effectsof event codability and of the ease of structural assembly on formulation points to a tight temporallink between the encoding of a relational, prelinguistic structure and its linguistic counterpart.Debates about structural processing in production often concern the abstractness of syntactic struc-tures (or syntactic plans; see Pickering & Ferreira, 2008, for a review), so the direction of these effectscan help distinguish between functional and abstract structural accounts of syntactic encoding.

These debates have so far been addressed in structural priming studies by testing the extent towhich repetition of structure from one sentence to another can be explained solely by priming ofconceptual–relational information (a functionalist perspective) and the extent to which structure-building procedures are independent of conceptual pressures (an abstract structural perspective;see Bock, 1982; Bock, Loebell, & Morey, 1992, for reviews). In principle, a sentence structure can bethe product of mapping operations that bind individual elements of a message representation (e.g.,characters in an event) to thematic roles (agents and patients) or it can be generated by structuralprocedures that are less sensitive to the identity of the characters filling those roles.

Examining the effects of structural primes on the timecourse of sentence formulation offers a newapproach to testing the nature of the dependencies between conceptual and linguistic structural pro-cesses. Functional accounts of syntax predict that the effect of structural primes should be limited topriming of thematic roles: an active prime should bias assignment of the agent to subject position anda passive prime should bias assignment of the patient to subject position (a form of prominence prim-ing; see Pickering & Ferreira, 2008). On this account, speakers in Experiment 2 should have quickly fix-ated and encoded the agent in the pictured event after hearing an active prime, and should havequickly fixated and encoded the patient after hearing a passive prime. This outcome would haveresembled accessibility effects obtained in Experiment 1 with lexical primes, supporting linear ratherthan hierarchical incrementality (but see Chang, Bock, & Goldberg, 2003, for priming of thematic rolesin a different structural alternation). Instead, structural priming in Experiment 2 favored encoding ofinformation about both characters in the event immediately after picture onset. The results show thatstructural procedures are concerned with expressing relational information rather than facilitating theassignment of a particular character to a particular structural slot, and is thus inconsistent withfunctional accounts of syntax.


Importantly, early effects of linguistic structure on formulation suggest an influence of linguisticprocesses on representations generated at the interface of message and sentence planning. A relatedquestion then is whether variables influencing encoding of linguistic relational information can alsoinfluence message content (or speakers’ perspective on an event). Recent work by Bunger,Papafragou, and Trueswell (in press) suggests that high similarity in the relational content of someconceptual representations and linguistic structures may indeed change what speakers express aboutan event after exposure to structural primes. There is little evidence to suggest that this might havebeen the case in the two experiments in this paper: the lexical primes in Experiment 1 and thestructural primes in Experiment 2 did not impact the heterogeneity of nouns and verbs chosen to referto the two characters or to describe the relationships between them.7 Thus overall, there is strongerevidence that primes shaped the way that the different increments of a message and sentence wereassembled rather than influencing what speakers said.

A compelling test of conceptual–linguistic priming is also afforded by cross-linguistic comparisonsof formulation. One long-standing, attractive hypothesis in the field is that differences in phrasal syn-tax across languages may support different patterns of incremental planning by requiring speakers toencode some types of information before others. Cross-linguistic studies therefore provide uniqueinsight into possible effects of linguistic structure on early sentence formulation and can help to iden-tify limits in the flexibility of production processes (Brown-Schmidt & Konopka, 2008; Christianson &Ferreira, 2005; Myachykov, Garrod, & Scheepers, 2009; Norcliffe, Konopka, Brown, & Levinson, 2013;Sauppe, Norcliffe, Konopka, van Valin, & Levinson, 2013).

4.4. Generalization to other types of messages

In our experiments, the non-relational and relational variables were uncorrelated. Namely, Eventcodability was not correlated with either Agent or Patient codability, and the content words selectedto refer to the two characters had no subcategorization preferences that either favored or disfavoredselection of active or passive syntax. In principle, however, properties of individual characters andproperties of events in normal production can operate independently as well as in concert to influencethe timecourse of formulation. For example, speakers’ choice of referential terms for individualcharacters may depend on their role in the event: selecting labels such as teacher and student fortwo characters is contingent on apprehension of the event structure. Identification of a character asa member of a particular profession can also influence the interpretation of the action performedby this character: e.g., verbs like shouting and directing may be better suited for sports fans and coachesrather than the other way around.

To test for interactions between these variables, it is important to first specify the level at whichinterdependence between non-relational and relational information may be observed. This requiresclarifying the sub-components of processes like event apprehension (or encoding of event gist), i.e.,explaining how speakers integrate various pieces of information about an event to create a represen-tation of its relational structure within the time it takes to process visual information during one ortwo fixations. On the one hand, research in memory and visual cognition has shown that peoplecan identify characters more quickly and accurately in coherent scenes than incoherent scenes (seeHenderson & Ferreira, 2004, for a review), supporting the idea of fast integration of non-relationaland relational information during construction of an event representation. On the other hand, encod-ing of event gist is more poorly defined in psycholinguistic work. For example, on Griffin and Bock’(2000) account, apprehension involves encoding enough information to specify the relationshipbetween two characters (chasing, kicking, etc.) and begin linguistic encoding, while on other accounts(e.g., Bunger et al., in press), identification of an event class (e.g., identifying an event as a motionevent) can also constitute encoding of event gist.

7 Effects of primes on sentence content were only observed in Experiment 2 with passive sentences: here, passive primesincreased rates of production of full passives and decreased rates of production of truncated passives. It is debatable, however, towhat extent expressing the agent in a passive sentence implies formulation of a different ‘‘message’’ than for a sentence where theagent is omitted.


Minimally, developing detailed models of event apprehension requires understanding how rela-tional information contributes to encoding of the non-relational content of an event, and vice versa.Hafri et al. (2012) recently showed that speakers can extract basic information about event structurein less than 100 ms from perceptual features of individual characters that are typically associatedwith ‘‘agenthood’’ (also see Bock et al., 2003; Dobel et al., 2007; Potter, 1976). Given the speed withwhich speakers can link visual information to event categories, the two experiments in this papersuggest that processing occurring within the first 400 ms of picture onset must be a multi-facetedprocess. Indeed, speakers did not fixate and continue fixating the character produced in subjectposition from picture onset until speech onset: plotting the timecourse of agent-directed fixationsin active sentences showed that, on average, speakers first fixated the agent and then the patientbefore 400 ms. Since it is possible to encode coarse-grained information about the event duringinitial fixations to the agent, this pattern suggests that fixating the second character served an addi-tional purpose before speakers redirected their gaze to the agent (the first-mentioned character). Thetime window argued to correspond to event apprehension by Griffin and Bock (2000) may thusencompass encoding of coarse-grained as well as finer-grained conceptual properties of an event;the extent to which these processes draw on non-relational and relational information remains tobe determined.

4.5. Conclusion

The timecourse of message formulation and sentence formulation can vary systematically fromcontext to context. Differences in the nature of the messages that speakers intend to communicateas well as moment-to-moment fluctuations in the speed of performing the necessary encodingoperations can create a bias for encoding either relational or non-relational information withpriority. Flexibility in the production system allows speakers to begin formulation by prioritizingencoding of information that is easy to process, and implicit choices of planning strategies haveconsequences for the timecourse of formulation from the generation of a message untilarticulation.

Of course, flexibility may be the rule rather than the exception for production outside of the lab asreal-life production contexts are undoubtably richer than in laboratory tasks. However, there mustalso be bounds on this flexibility. At the extreme, radical linear incrementality is unlikely to accountfor formulation of sentences with a complex conceptual structure because some form of conceptualguidance is necessary for speakers to structure sentences around the ‘‘thought’’ that they want to com-municate. Hierarchical incrementality is also unlikely to mediate construction of simpler phrases (e.g.,conjuncts), where word order may reflect differences in the order of word activation (axe and saw orsaw and axe) or common usage (king and queen but not queen and king). Thus as in studies examiningcontext effects on various aspects of on-line processing (e.g., use of common ground in conversationalexchanges; Brown-Schmidt & Konopka, 2011), an emphasis on flexibility requires further specificationof how and when different variables shape formulation.

Author note

These experiments were first presented at the 18th meeting of the AMLaP conference in 2011. Wethank Moniek Schaars for invaluable help with data collection and processing, and Katrien Scheibe andSamantha Hoogen for assistance with data collection.

Appendix A

Target events presented in Experiment 1 (the agent, patient, and unrelated prime sentences inDutch and English respectively are listed in parentheses).


1. Bulldozer tearing down building(De machine is lawaaiig / Het paleis staat op de top van de berg / De balletdansers dansen)(The machine is noisy / The palace is at the top of the hill / The ballet dancers are dancing)

2. Tree crushing man(Het hout is klaar om vervoert te worden / De wandelaar is aan het klimmen / De olifant drinkt)(The wood is ready to be shipped / The hiker is climbing / The elephant is drinking)

3. Bee stinging man(De honing is in de aanbieding / De opa is aan het lezen / De robot loopt)(The honey is on sale / The grandfather is reading / The robot is walking)

4. Man shooting woman(De tiener rolschaatst / Het meisje slaapt / De bloemen zijn verwelkt)(The teenager is rollerskating / The girl is sleeping / The flowers are wilting)

5. Baseball hitting boy(De shuttle ligt in het gras / De man ligt in de hangmat / De telefoon rinkelt)(The shuttle is lying in the grass / The man is lying in a hammock / The phone is ringing)

6. Dog chasing mailman(De wolf huilt naar de maan / De verkoper wacht buiten / De paraplu waait weg)(The wolf is howling at the moon / The salesman is waiting outside / The umbrella is flying away)

7. Journalist interviewing actor(De schrijver denkt / De muzikant buigt / Het hert verstopt zich achter de struik)(The writer is thinking / The musician is bowing / The deer is hiding behind a bush)

8. Boxer hitting cheerleader(De gladiator rust uit / De fans juichen / Het zaad komt uit)(The gladiator is resting / The fans are cheering / The seed is sprouting)

9. Snake scaring girl(De hagedis zit op een steen / Het kind is aan het schommelen / De bal rolt van de berg)(The lizard is sitting on a stone / The child is swinging / The ball is rolling down the hill)

10. Masseuse massaging man(De kapper werkt / De jongen niest / De hut staat in brand)(The hairdresser is working / The boy is sneezing / The hut is burning)

11. Shark swallowing swimmer(De goudvis zwemt in de kom / De duikbril hangt aan de hak / De wieltjes draaien)(The goldfish is swimming in a bowl / The goggles are hanging on a hook / The wheels are turning)

12. Photographer filming model(De schilder is aan het werk / De paspoppen staan in de hoek / De eieren bakken in de koekenpan)(The painter is working / The mannequins are standing in the corner / The eggs are frying)

13. Policeman stopping driver(De detective is aan het roken / De fietser rijdt een heuvel op / De vlinder zit op een bord)(The detective is smoking / The biker is riding up the hill / The butterfly is sitting on a plate)

14. Lighting striking church(De regen valt neer / De oude bel luidt / Het konijn komt uit de hoed)(The rain is pouring down / The old bell is ringing / The rabbit is coming out of the hat)

15. Policeman arresting man(De bewaker staat bij het hek / De jongen rilt / De vulkaan barst uit)(The guard is standing by the gate / The boy is shivering / The volcano is erupting)

16. Bomb hitting ship(Het pistool ligt in de lade / De kano’s liggen op het strand / De farao zwaait)(The gun is lying in the drawer / The canoes are lying on the beach / The pharaoh is waving)

17. Fireman saving child(De soldaat sloopt over de balk / De baby kruipt / De beer gromt)(The soldier is walking on a log / The baby is crawling / The bear is growling)

18. Bird pulling worm out of the ground(De pinguin schaatst / De kever zit in de pot / De serveerster struikelt)(The penguin is skating / The beetle is sitting in a jar / The waitress tripped)

19. Army attacking castle(De tank is aan het schieten / De muur brokkelt af / Het water kookt in de ketel)(The tank is firing / The wall is crumbling / The water is boiling in the kettle)

20. Alarm waking up boy(De luidspreker staat op de venster bank / Het meisje is touwtje aan het springen / Het hek staat open)(The loudspeaker is standing on the windowsill / The girl is skipping rope / The gate is open)

21. Horse kicking cow(Het zadel hangt op het hek / De melkfles staat bij de deur / De gevangene zit achter de tralies)(The saddle is hanging on the fence / The bottle of milk is at the door / The prisoner is sitting behind bars)


22. Judge giving tennis player a medal(De scheidsrechter wijst / De hardloper rekt zijn spieren / De vaas valt van de tafel)(The referee is pointing / The runner is stretching / The vase is falling off the table)

23. Thief stealing painting(De spion is aan het gluren / Het beeld staat in een doos / De ballon knapt)(The spy is spying / The statue is standing in a box / The balloon is bursting)

24. Ambulance hitting woman(Het ziekenhuisbed is leeg / Het dienstmeisje is aan het stofzuigen / De ventilator staat aan)(The hospital bed is empty / The maid is vacuuming / The fan is turned on)

25. Bishop crowning king(De monnik is aan het bidden / De president spreekt / De sneeuwpop smelt)(The monk is praying / The president is speaking / The snowman is melting)

26. Helicopter pulling diver out of the water(Het vliegtuig vertrekt / De matroos drinkt / De pannen hangen in een rij)(The airplane is taking off / The sailor is drinking / The pots are hanging in a row)

27. Sprinkler splashing old woman(De gieter lekt / De heks zit op een bezem stil / De cd is gebroken)(The watering can is leaking / The witch is riding on a broom / The cd is broken)

28. Truck towing car(De sneeuwschuiver wacht voor het stoplicht / De motorfiets staat tegen de lantaarnpaal / De dwerg lacht)(The snowplow is standing at the stoplight / The motorcycle is leaning against the lamp post / The dwarf islaughing)

29. Cat catching mouse(De tijger springt door een hoepel / De kaas ligt in de koelkast / De man is zich aan het scheren)(The tiger is jumping through a hoop / The cheese is in the fridge / The man is shaving)

30. Nurse vaccinating patient(De chirurg is moe / Het slachtoffer is bewusteloos / De kleren drogen aan de waslijn)(The surgeon is tired / The victim is unconscious / The clothes are drying on a line)

Appendix B

Targets events presented in Experiment 2 (the transitive and intransitive prime sentences in Dutchand English are listed in parentheses).

1. Bulldozer tearing down building(De alien trekt de astronaut mee / Twee mensen roeien)(The alien is pulling the astronaut / Two people are rowing)

2. Thief stealing painting(De professor feliciteert de student / /De eieren bakken in de koekenpan)(The professor is congratulating the student / The eggs are frying in the pan)

3. Bee stinging man(De steen heeft het raam gebroken/ De kleren drogen)(The rock is breaking the window / The clothes are drying)

4. Man shooting woman(De ooievaar brengt de baby / De bloemen zijn verwelkt)(The stork is bringing the baby / The flowers are wilting)

5. Robot smashing computer(Het dienstmeisje gooit het vuilnis weg / Twee mensen trouwen)(The maid is throwing out the garbage / Two people are getting married)

6. Dog chasing mailman(De technicus maakt de tv / Enkele munten vallen uit een zak)(The technician is fixing the tv / The coins are falling out of the bag)

7. Journalist interviewing actor(De kikker vangt de vlieg / De beer en de pinguin dansen)(The frog is catching the fly / The bear and the penguin are dancing)

8. Boxer hitting cheerleader(De moeder kleedt de jongen aan / De monniken bidden)(The mother is dressing the boy / Two monks are praying)

9. Snake scaring girl(De coach omhelst de tennister / De kaas en de wijn zijn in de aanbieding)(The coach is hugging the tennis player / The cheese and the wine are on sale)

10. Masseuse massaging man

(continued on next page)


(De eendjes volgen moeder eend/ De herten verstoppen zich achter de struiken)(The ducklings are following the mother duck / The deer are hiding behind the bushes)

11. Shark swallowing man(De Nederlandse speler maakt een doelpunt / De bellen luiden)(The Dutch player is scoring a goal / The bells are ringing)

12. Camera man filming model(De trein verplettert de bus / De olievaten lekken)(The train is crushing the bus / The oil cans are leaking)

13. Policeman stopping car(De leeuw verslindt de zebra / De hagedissen zitten op een steen)(The lion is devouring the zebra / The lizards are sitting on a stone)

14. Lightning striking church(De kok versiert de taart / The soldaten marcheren)(The cook is decorating the cake / The soldiers are marching)

15. Girl tickling boy(De dirigent leidt het orkest / De goudvissen zwemmen in de kom)(The conductor is conducting the orchestra / The goldfish are swimming in a bowl)

16. Bomb hitting ship(De kerstman versleept de kerstboom / De schapen grazen)(Santa Claus is dragging the Christmas tree / The sheep are grazing)

17. Fireman saving boy(De hond leidt de blinde man / De blinde man en de hond wachten bij de kruising)(The dog is leading the blind man / The blind man and the dog are waiting at the crossing)

18. Bird pulling worm out of the ground(De burgemeester onthult het standbeeld / Het stel is aan het rolschaatsen)(The mayor is unveiling the statue / The couple are rollerskating)

19. Army attacking castle(Het meisje duwt de jongen op de slee / De cadeautjes liggen onder de kerstboom)(The girl is pushing the boy on the sled / The presents are under the Christmas tree)

20. Alarm waking up boy(De wind blaast de papieren weg / De zadels hangen op het hek)(The wind is blowing the papers away / The saddles are hanging on a fence)

21. Horse kicking cow(De bodyguard trekt de president weg / De bewakers staan voor de poort)(The bodyguard is pulling the president away / Two guards are standing in front of the gate)

22. Policeman arresting man(De baas stelt de secretaresse voor / De ballonnen knappen)(The boss is introducing the secretary / The balloons are bursting)

23. Elephant lifting clown(De piraat begraaft de schat / Het team is aan het praten)(The pirate is burying the treasure / The team is talking)

24. Ambulance hitting woman(De padvinder roostert het varken / Het oudere paar houdt elkaars handen vast)(The boyscout is roasting the pig / The older couple are holding hands)

25. Bishop crowing prince(De lawine treft de skiers / De sneeuwpoppen smelten)(The avalanche is hitting the skiers / The snowmen are melting)

26. Helicopter pulling diver out of the water(De juwelier onderzoekt de diamanten / Het koor zingt)(The jeweler is examining the diamonds / The choir is singing)

27. Sprinkler splashing old woman(De gastvrouw laat de man binnen / De fans juichen)(The hostess is inviting the man in / The fans are cheering)

28. Nurse carrying baby(De rivier breekt de dam door / De mannen zijn aan het skydiving)(The river is breaking the dam / The men are skydiving)

29.Cat catching mouse(De oude ober schopt de man van de trap / Twee mensen lopen onder een paraplu)(The old waiter is kicking the man down the stairs / Two people are walking under an umbrella)

30. Truck towing car(De hond likt in het gezicht van de jongen / De dwergen lachen)(The dog is licking the boy’s face / The dwarfs are laughing)


References

Altmann, L. J. P., & Kemper, S. (2006). Effects of age, animacy and activation order on sentence production. Language andCognitive Processes, 21, 322–354.

Baayen, R. H., Davidson, D. J., & Bates, D. M. (2008). Mixed-effects modeling with crossed random effects for subjects and items.Journal of Memory and Language, 59, 390–412.

Barr, D. J. (2008). Analyzing ‘visual world’ eyetracking data using multilevel logistic regression. Journal of Memory and Language,59, 457–474.

Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing: Keep itmaximal. Journal of Memory and Language, 68, 255–278.

Bates, E., & MacWhinney, B. (1982). Functionalist approaches to grammar. In E. Wanner & L. R. Gleitman (Eds.), Languageacquisition: The state of the art (pp. 173–218). Cambridge: Cambridge University Press.

Bock, J. K. (1982). Toward a cognitive psychology of syntax: Information processing contributions to sentence formulation.Psychological Review, 89, 1–47.

Bock, J. K. (1986a). Syntactic persistence in language production. Cognitive Psychology, 18, 355–387.Bock, J. K. (1986b). Meaning, sound, and syntax: Lexical priming in sentence production. Journal of Experimental Psychology:

Learning, Memory, and Cognition, 12, 575–586.Bock, J. K. (1987a). Coordinating words and syntax in speech plans. In A. Ellis (Ed.). Progress in the psychology of language (Vol. 3).

London: Erlbaum.Bock, J. K. (1987b). An effect of the accessibility of word forms on sentence structures. Journal of Memory and Language, 26,

119–137.Bock, J. K., & Griffin, Z. M. (2000). The persistence of structural priming: Transient activation or implicit learning? Journal of

Experimental Psychology: General, 129, 177–192.Bock, J. K., & Irwin, D. (1980). Syntactic effects of information availability in sentence production. Journal of Verbal Learning and

Verbal Behavior, 19, 467–484.Bock, J. K., Irwin, D. E., & Davidson, D. J. J. (2004). Putting first things first. In F. Ferreira & M. Henderson (Eds.), The integration of

language, vision, and action: Eye movements and the visual world (pp. 249–278). New York: Psychology Press.Bock, J. K., Irwin, D. E., Davidson, D. J., & Levelt, W. J. M. (2003). Minding the clock. Journal of Memory and Language, 48, 653–685.Bock, J. K., Loebell, H., & Morey, R. (1992). From conceptual roles to structural relations: Bridging the syntactic cleft. Psychological

Review, 99, 150–171.Bock, J. K., & Warren, R. K. (1985). Conceptual accessibility and syntactic structure in sentence formulation. Cognition, 21, 47–67.Branigan, H. P., Pickering, M. J., & Tanaka, M. (2008). Contributions of animacy to grammatical function assignment and word

order during production. Lingua, 118, 172–189.Brown-Schmidt, S., & Konopka, A. E. (2008). Little houses and casas pequeñas: Message formulation and syntactic form in

unscripted speech with speakers of English and Spanish. Cognition, 109, 274–280.Brown-Schmidt, S., & Konopka, A. E. (2011). Experimental approaches to referential domains and the on-line processing of

referring expressions in unscripted conversation. Information, 2, 302–326.Brown-Schmidt, S., & Tanenhaus, M. K. (2006). Watching the eyes when talking about size: An investigation of message

formulation and utterance planning. Journal of Memory and Language, 54, 592–609.Bunger, A., Papafragou, A., & Trueswell, J. (2013). Event structure influences language production: Evidence from structural

priming in motion event description. Journal of Memory and Language, 69(299), 323.Chang, F., Bock, K., & Goldberg, A. (2003). Can thematic roles leaves traces of their places? Cognition, 90, 29–49.Chang, F., Dell, G. S., & Bock, K. (2006). Becoming syntactic. Psychological Review, 113, 234–272.Christianson, K., & Ferreira, F. (2005). Conceptual accessibility and sentence production in a free word order language (Odawa).

Cognition, 98, 105–135.Dobel, C., Gumnior, H., Bolte, J., & Zwitserlood, P. (2007). Describing scenes hardly seen. Acta Psychologica, 125, 129–143.Ferreira, F. (1994). Choice of passive voice is affected by verb type and animacy. Journal of Memory and Language, 33(6), 715–736.Ferreira, V. S. (1996). Is it better to give than to donate? Syntactic flexibility in language production. Journal of Memory and

Language, 35, 724–755.Ferreira, V. S., & Bock, K. (2006). The functions of structural priming. Language and Cognitive Processes, 21, 1011–1029.Ferreira, F., & Henderson, J. M. (1998). Linearization strategies during language production. Memory and Cognition, 26, 88–96.Ferreira, F., & Swets, B. (2002). How incremental is language production? Evidence from the production of utterances requiring

the computation of arithmetic sums. Journal of Memory and Language, 46, 57–84.Fisher, C. (2002). The role of abstract syntactic knowledge in language acquisition: A reply to Tomasello (2000). Cognition, 82,

259–278.Fisher, C., Gleitman, H., & Gleitman, L. R. (1991). On the semantic content of subcategorization frames. Cognitive Psychology, 23,

331–392.Gleitman, L., January, D., Nappa, R., & Trueswell, J. C. (2007). On the give and take between event apprehension and utterance

formulation. Journal of Memory and Language, 57, 544–569.Griffin, Z. M., & Bock, K. (2000). What the eyes say about speaking. Psychological Science, 11, 274–279.Griffin, Z. M. (2004). Why look? Reasons for speech-related eye movements. In F. Ferreira & M. Henderson (Eds.), The integration

of language, vision, and action: Eye movements and the visual world (pp. 191–211). New York: Psychology Press.Hafri, A., Papafragou, A., & Trueswell, J. C. (2012). Getting the gist of events: Recognition of two-participant actions from brief

displays. Journal of Experimental Psychology: General. http://dx.doi.org/10.1037/a0030045.Hartsuiker, R. J., Bernolet, S., Schoonbaert, S., Speybroeck, S., & Vanderelst, D. (2008). Syntactic priming persists while the lexical

boost decays: Evidence from written and spoken dialogue. Journal of Memory and Language, 58, 214–238.Henderson, J. M., & Ferreira, F. (2004). Scene perception for psycholinguistics. In F. Ferreira & M. Henderson (Eds.), The

integration of language, vision, and action: Eye movements and the visual world (pp. 1–58). New York: Psychology Press.Henderson, J. M., Malcolm, G. L., & Schandl, C. (2009). Searching in the dark: Cognitive relevance drives attention in real-world

scenes. Psychonomic Bulletin and Review, 16, 850–856.

http://refhub.elsevier.com/S0010-0285(14)00027-9/h0005




























































http://dx.doi.org/10.1037/a0030045








Jaeger, T. F. (2008). Categorical data analysis: Away from ANOVAs (transformation or not) and towards logit mixed models.Journal of Memory and Language, 59, 434–446.

Kempen, G., & Hoenkamp, E. (1987). An incremental procedural grammar of sentence formulation. Cognitive Science, 11,201–258.

Konopka, A. E. (2012). Planning ahead: How recent experience with structures and words changes the scope of linguisticplanning. Journal of Memory and Language, 66, 143–162.

Kuchinsky, S. E. (2009). From seeing to saying: Perceiving, planning, producing. Unpublished doctoral dissertation, University ofIllinois at Urbana-Champaign.

Kuchinsky, S. E., & Bock, K. (2010). From seeing to saying: Perceiving, planning, producing. Paper presented at the 23rd meeting ofthe CUNY human sentence processing conference, New York, NY.

Kuchinsky, S. E., Bock, K., & Irwin, D. E. (2011). Reversing the hands of time: Changing the mapping from seeing to saying. Journalof Experimental Psychology: Learning, Memory, and Cognition, 37, 748–756.

Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press.McDonald, J. L., Bock, K., & Kelly, M. H. (1993). Word and world order: Semantic, phonological, and metrical determinants of

serial position. Cognitive Psychology, 25, 188–230.Meyer, A. S., & Lethaus, F. (2004). The use of eye tracking in studies of sentence generation. In F. Ferreira & M. Henderson (Eds.),

The integration of language, vision, and action: Eye movements and the visual world (pp. 191–211). New York: PsychologyPress.

Myachykov, A., Garrod, S., & Scheepers, C. (2009). Attention and syntax in sentence production: A critical review. Discours,4(special issue: Linearization and Segmentation in Discourse). http://dx.doi.org/10.4000/discours.7594.

Myachykov, A., Garrod, S., & Scheepers, C. (2012). Determinants of structure choice in visually situated sentence production.Acta Psychologica, 141, 304–315.

Norcliffe, E., Konopka, A. E., Brown, P., & Levinson, S. (2013). Word order affects the timecourse of sentence formulation inTzeltal. Paper presented at the 26th meeting of the CUNY human sentence processing conference, Columbia, SC, USA.

Osgood, C. E., & Bock, J. K. (1977). Salience and sentencing: Some production principles. In S. Rosenberg (Ed.), Sentenceproduction: Developments in research and theory (pp. 89–140). Hillsdale, NJ: Erlbaum.

Pickering, M. J., & Branigan, H. P. (1998). The representation of verbs: Evidence from syntactic priming in language production.Journal of Memory and Language, 39, 633–651.

Pickering, M. J., & Ferreira, V. S. (2008). Structural priming: A critical review. Psychological Bulletin, 134, 427–459.Potter, M. C. (1976). Short-term conceptual memory for pictures. Journal of Experimental Psychology: Human Learning and

Memory, 2, 509–522.Prat-Sala, M., & Branigan, H. P. (2000). Discourse constraints on syntactic processing in language production: A cross-linguistic

study in English and Spanish. Journal of Memory and Language, 42, 168–182.Sauppe, S., Norcliffe, E., Konopka, A. E., van Valin, R. D., & Levinson, S. (2013). Dependencies first: Eye-tracking evidence from

sentence production in Tagalog. Paper to be presented at the 35th meeting of the Cognitive Science Society, Berlin, Germany.Smith, M., & Wheeldon, L. (2001). Syntactic priming in spoken sentence production – An online study. Cognition, 78, 123–164.Tomasello, M. (2000). Do young children have adult syntactic competence? Cognition, 74, 209–253.Tomlin, R. (1995). Focal attention, voice, and word order: An experimental, cross-linguistic study. In M. Noonan & P. Downing

(Eds.), Word order in discourse (pp. 521–558). Amsterdam: John Benjamins.Tomlin, R. (1997). Mapping conceptual representations into linguistic representations: The role of attention in grammar. In J.

Nuyts & E. Pederson (Eds.), Language and conceptualization (pp. 162–189). Cambridge: Cambridge University Press.Van de Velde, M., Meyer, A. S., & Konopka, A. E. (2014). Message formulation and structural assembly: Describing ‘‘easy’’ and

‘‘hard’’ events with preferred and dispreferred syntactic structures. Journal of Memory and Language, 71, 124–144.Wagner, V., Jescheniak, J. D., & Schriefers, H. (2010). On the flexibility of grammatical advance planning during sentence

production: Effects of cognitive load on multiple lexical access. Journal of Experimental Psychology: Learning, Memory, andCognition, 36, 423–440.

Wheeldon, L. (2013). Producing spoken sentences: The scope of incremental planning. In P. Perrier and P. Lang Verlag (Eds.),Cognitive and physical models of speech production, speech perception, and production-perception integration.















http://dx.doi.org/10.4000/discours.7594























priming sentence planning - mpg.pure

Documents